CUHKSZ’s paired hand-gripper dataset of 6,189 episodes hits 70% real-robot success on Galaxea R1 Lite for spatial pick-and-place.
UTS trains a UR5e to adapt demonstrated skills around moving obstacles; hardware success hits 90% versus 10% for the prior GP baseline.
UCL and PolyU researchers spoof Franka telemetry through LD_PRELOAD, bypassing SROS 2 and beating an AI detector 87% of the time.
EPFL and HKUST open-source a ROS 2 swarm stack on OmniNxt hardware, with a six-drone swap and a four-drone obstacle run.
The Guangxi factory opened 12 September with a 10-minute takt, Cruzr destacking, 2,000 screws per chassis, and a 65 m² rack for 112 finished units.
Munro Live’s G1 joint teardown finds ~15:1 planetary gearing, knee heat pipes, remote ankle motors, and CNC parts built for iteration.
The Paris humanoid, built to haul tires, panels, and carts, now has 12 enterprise buyers; Renault still plans 350 units in 18 months.
SJTU’s ray-scaling layer hits 100% step safety on π0.5 and Fast-WAM, then keeps a real ALOHA arm on a magnetic-ball track.
Hero Esports staged ten full-size T800s in a Riyadh arena on Sept. 9 and sketched an eight-city robot sports league.
A dual-camera latent world model hits 92% on reach and 72.5% on angled-reach, beating single-view V-JEPA 2-AC.
A 31-DoF lab humanoid with independently actuated RGB-D heads lifts visible-reachable coverage from 38% to 97%.
A 33,985-parameter probe on VLA-JEPA states hits 85.68 pooled AUROC and transfers to PIPER X, SO-101, and Franka.
AGIBOT and Xi’an Jiaotong’s DCT-conditioned expert gains 9.3 points on LIBERO-Plus and lifts six A2 humanoid tasks.
A 21-member electric truss on four wheels traverses 2.11 wheel radii on hardware by reusing four contact primitives.
GMO AIR’s Tokyo van carries engineers, parts, and a replacement humanoid so a broken unit can be swapped on site.
A month after stealth on Unitree G1s, Nucleus deployed a wheeled dual-arm with 10 kg per arm and 5–9 hour uptime.
A 17-family, 1,000-scene benchmark puts MLLMs in a simulated humanoid and scores whether the catch actually meets the object.
Fudan’s CoRL 2026 paper puts a 20-DoF humanoid on overhead bars with passive wrist hooks and a residual-keyframe RL stack.
KU Leuven and Meituan split a VLM agent from a stateless VLA. Grounded grasp/place/move hints lift LIBERO-Mem relaxed success to 63%.
HD Hyundai Robotics puts 13 billion won into AIDIN's force-torque sensors and plans a five-finger hand for shipbuilding humanoids.
Fudan and TARS train one G1 policy that denoises corrupted depth instead of switching to a blind controller. CoRL 2026.
Zhejiang’s InReal hand uses Graph-Mimic and a key-velocity loudness model. Real F1 hits 0.80–0.95; untrained listeners don’t prefer the human.
A native world-action model, trained from scratch, lifts G1-OP success from 17.1% to 44.1% across 100 unseen tasks. Code is coming.
Yonsei and RLWRLD robotize egocentric clips into 142 million frames. Pretraining lifts real ALLEX completion from 51.5% to 80.3%.
SFU and Penn replace π0.5's flow-matching loop with one cIMLE pass. 98% on LIBERO, smoother Franka motion, IROS 2026.
KIMM's 170 cm, 29-DoF homegrown humanoid shows mask dance and calisthenics in Seoul. V1 is dated April 2027.
HIT's memory-as-plans stack hits 83.3% on RMBench and 78% on a Franka that must recall covered buttons and numbers.
Santa Clara’s Maven Robotics exits stealth with vacuum-grip mixed palletizing: 10 mph, 30 kg, up to eight robots, 16-hour days.
ETH Zurich estimates a task Jacobian on the fly. A 17-DoF tendon hand writes on paper at 0.64 mm error with no sim, no demos.
Zhejiang and NUS train a perceptive motion tracker that adapts flat-ground references to real terrain without matched demos.
SEED-UMI puts one 20-DoF glove on both a person and a Wuji hand; paired replay hits 70% success on five contact-rich tasks.
A memory-prediction-action VLA beats π0.5 on LIBERO-Plus and a Galaxea R1 Lite, with 25–50% of the baseline’s training epochs.
Agile Robots is building its 174 cm humanoid in Fürstenfeldbruck, with a 21-DoF hand and DeepMind insertion numbers attached.
Fudan, Midea, and CMU release 2,000+ hours of folding data across 10+ robots, plus a FoldChallenge leaderboard.
Berkeley and Sharpa ground 1,500+ Seedance clips in Isaac Gym, then run a Sharpa Wave hand at 27/40 on unseen plans.
CMU and Bosch VLMs compose task-axis controllers on a UR5e. Zero-shot success is 73.9% across 14 tasks; one round of feedback lifts that to 90.7%.
Jilin and JD post-train OpenVLA-OFT in a world model, then raise G1 success from 67.5% to 80% by scoring fake futures.
HKU and HKUST Guangzhou train fixture-free in-hand assembly in sim; a 22-DoF hand mates two parts with a single camera, 18/20 on bottle and syringe.
A two-pass Transformer predicts the action chunk and the future latent together. 83% in sim, 66.9% across 630 real episodes, 13.2 ms of model time.
NUS Show Lab turns MV_LEFT into a Franka step. Gemini-3.1 Pro hits 89% zero-shot; a 2B Qwen fine-tune is close behind.
Saxion researchers anneal PD stiffness from 60 to 40 so a Boston Dynamics Spot can tip 90 degrees onto its front legs; 10 seeds transfer to hardware.
UMass and MIT decouple VR head motion from robot actuation, cutting drop-recovery time to 19.7 s across 10 operators.
IEEE Spectrum surveys T-Rex’s 65% touch-reactive VLA, a 3,000-hour multi-sensor set, and NeoteAI’s 30,000-hour tactile world-action model.
A 6B whole-body VLA trained on ~2,500 hours of real-robot data, covering tabletop work and laundry-to-washer locomotion.
USC post-trains a frozen SONIC tracker with LoRA and a visual extractor; a G1 parkours, reorients cubes, and dodges balls zero-shot from sim.
HKUST Guangzhou and IIT turn pedal takeovers into a risk head, hitting 96.7% on hard grasps and cutting operator steps 84%.
D-Robotics says TCL’s hey AiMe, Vbot’s SuperDog, and the xLean TR1 all run its Sunrise AI silicon on the Berlin floor.
Peking and BAAI’s CoRL 2026 model gates tactile input on contact and hits 71% success on six real SharpaWave-hand tasks.
Cornell’s Proxy Policy Steering lifts π0.5 by 53 points on average and turns coffee brewing from 0/10 to 5/10 while keeping recovery behaviors.
A CoRL 2026 VLA predicts 29-DoF G1 actions from language and RGB, trained only in sim, then zero-shot through a 30-meter office.
Montreal automation firm’s new lab, led by Jimmy Li and advised by Cohere’s Joelle Pineau, will post-train robot manipulation on live plant data.
A wheeled-legged guide robot launched at Beijing’s rehab expo claims 99% indoor obstacle avoidance and a Hangzhou street pilot.
Agility’s Digit Redecorates clip shows the warehouse biped shoving an orange sofa and shuffling mixed clutter, the same grip variety it uses on totes.
A PolyU dressing policy learned on still arms, then tracks a moving sleeve with point-cloud registration. Nine people, three garments, 89% insert.
IIT fine-tunes GR00T N1.5 with a second of fingertip history plus compliance. Full system: 93% handover success, 9 of 10 people ranked it first.
Wayne State’s FailureSpot times VLA failure onset on LIBERO-10. Weak labels from action chunks plus 15% dense annotation beat trajectory-wide alarms.
Galbot and Tsinghua’s ECCV 2026 system generates co-speech gestures on a Unitree G1 with BrainCo hands at about 120 FPS, then filters collisions.
A Budapest study splits Jacobian from tendons on the Shadow Hand and a biomechatronic copy. Anatomy helps the thumb’s kinematics and wrecks its actuation.
BGR’s IFA walk: PrimeBOT Q1 kung fu, Booster K1 in a Messi kit from $5,999, EngineAI T800 at 450 N·m. PrimeBOT also took IFA Innovation Awards.
A 2,178-scenario benchmark built from real VLA failures. Near-perfect LIBERO scores fall more than 50% when the policy has to recover.
EMNLP 2026 benchmark with 527K trajectories. At difficulty L5, π0.5, X-VLA, GO-1, and RDT all sit under 25% success.
OpenMind’s onboard G1 stack maps wave, stop, and ‘I’m on a phone’ to conservative greeter behaviors in 25 ms on Jetson Thor, no cloud in the loop.
A Flexiv world-action model caches predicted tactile contact, then a 32.9 Hz expert patches the unexecuted suffix. Success rises from 22% to 64%.
Unitree demos UnifoLM-X2-1.0, a world-action model that lets a G1 punch, kick, and slip a padded human partner with no VR or gamepad in the loop.
USTC’s VLA-Precision posts 98.3% mean success on nine chemistry tasks across four robot setups, with 45.8 minutes of online RL per task.
XPeng commissions a Guangzhou IRON line with >80% core automation; a finished robot walks off under its own power. Mass production still set for year-end.
A new VLG stack keeps a hand-agnostic grasp policy and bolts on spatial, cognitive, and temporal priors. Real clutter: 83.3% over 510 trials.
A training-free UAV–UGV stack shares a bird’s-eye map. On 100 CARLA-Air episodes it hits 77% joint success, +27 points over the weaker agent.
AgiBot and Wuhan University replace binary contact gating with continuous per-foot FK reliability. Real walking ATE drops 70.8% on A3 Ultra.
CASIA, Tsinghua, and NUS post GIFT, a feature-training method that lifts LIBERO-Plus and real-arm success without changing how actions are decoded.
RoboCurve ran OpenAI’s GPT-6 Astra on dual I2RT YAM arms. Gross pick-and-place is cheap and fast. Millimeter insertion is still 2 of 20.
HINT injects tracked intent into π0.5 and Wall-OSS-0.5 with no extra trainable weights. Dual-arm spelling full-task success jumps from 13.3% to 86.7%.
ZINOVA and LimX mount TRON 2 dual-arm torsos on orange industrial booms for a scaled tilt-up demo: boards, nail guns, rebar. Teleop, not a job site yet.
PAMoR conditions real-time Unitree G1 motion on kinematics-measured valence and arousal. Raters hit 0.38 top-1, near acted human bodies at 0.44.
A Riyadh pipeline prints unactuated grippers from a mesh, pose, and six-axis robot. Four tools pass digital gates; physical tests pending.
A real-to-sim plus VLM-preference pipeline matches human pairwise judgments 82.9% of the time on LIBERO and 91.9% on calibrated hardware tasks.
Rice and NVIDIA retrieve human demos from internet video using actor-centered 3D hand motion. Downstream dexterous PPO jumps on harder VTDexManip tasks.
Stay Seated learns omnidirectional seated locomotion on a passive five-caster chair. The G1 policy uses only proprioception and transfers zero-shot.
SNU, ETH, and RAI’s RA-L paper reweights radar Doppler toward weak LiDAR directions. On three degenerate sequences, ATE drops 86% versus fixed fusion.
Tuya Smart debuts Doova at IFA: LiDAR, four mics, skeletal vision, a 60-second family video call, and a 10.1-inch face. No price yet.
WORLD 2026 Theme 3 gives researchers 14 tasks of expert demos, 1,024 successful rollouts, 1,369 failures, and human takeovers on G2.
AIST and Osaka’s 7-DoF gripper mixes a vacuum jamming palm with Fin-Ray fingertips and is accepted to IEEE TASE.
SmoothRL only backprops through the action chunk the Astribot S1 actually executed, lifting throw success from 39% to 94%.
Pathak, Gupta, and collaborators open-source Bridge: 21 DoF, about $1,500, trained by iterating morphology against human motion tracking.
UNC and MERL add evidence-gated training so π0.5 ignores junk sensors. On Kinova arms, distractor success jumps from 30% to 85%.
A Deep Robotics M20S estimates contact from residual torque, then a VLA and whole-body compensator share that signal for wiping and door work.
Berkeley, CMU, and Stanford train a G1 stop policy plus two stoppability critics, with 96.4% OOD stops and hardware clips that still fall.
Hypershell’s Halo full-leg exo coordinates hip and knee motors at 1,490 W peak. Origin Edition pre-orders start at $2,299.
Arizona State researchers run Koopman MPC on 3- and 5-segment cable-driven arms at 0.6 m/s, with 400 g tip load and a 7 N shove.
Fudan’s diffusion navigator hits 89.7% sim success and a 1 km outdoor follow on Unitree Go1, accepted to IROS 2026.
Queen Mary’s multi-vine eversion robot grows a tool channel from the outside and turns by locking one vine, aimed at EndoTheranostics.
Tokyo’s Jizai opens early access to Palmimo DevKit at ¥598,000, with a 21-DoF tabletop body, Raspberry Pi 5, and an open Python SDK.
Tsinghua and BAAI schedule world-model rollouts at contact-heavy states. On a Galaxea R1 Lite, π0 success rises 17.5 points.
PrimeBot and Peking University open 531.7 robot hours plus ~1,000 UMI hours, then push clothes-folding success from 58% to 93% with corrections.
ADAPT maps language to joint actions with a diffusion prior plus residual RL, deployed on a Unitree G1 at 50 Hz.
Lumos Robotics’ MOS 2 is a 1.65 m wheeled dual-arm with 50 kg max payload, already on Mitsubishi Electric PLC lines.
HKUST(GZ) and Noitom’s UMR learns dense surface correspondence so G1, H2, and T800 share one retargeter.
D-Robotics’ WM-LOCO co-trains a world model with PPO; a Unitree G1 hits 93.3% on stones, stairs, and a gap.
X Square Robot’s TwinDEX pairs a 3-finger, 9-DoF wearable with a matching robot hand for robot-free dexterous data.
Suzhou’s Zeroth launched Bridge, an 88 cm, 13 kg humanoid, with OpenBridge and Geek Edition bookings from Sept 2.
AI² Robotics parked AlphaBot 2 behind a Central bar on Aug 31. Allan Zeman says tablet orders and pours start about two weeks later.
HKU and Oxford’s ConsJump policy jumps a Unitree Aliengo through a narrow gate at 2.5 m/s with 0.44 s in the air.
ETH Zurich’s PM-01 policy jumps onto bars, brachiates at 0.5 m/s, and lands in 14 of 15 hardware trials using a head-mounted solid-state lidar.
NTU’s Facet-0 predicts wrist wrench with each action chunk and hits 82% on five computer-assembly tasks versus 15% for the best baseline.
At WRC 2026, Galaxea premiered Nexo: 30 DoF, 20 kg dual-arm payload, 8-hour runtime, aimed at packing and assembly rather than flips.
Galbot will take pre-orders for ET1, its 1,230 mm, 30 kg biped on AstraBrain-WBC, starting 3 September. No price in the materials fetched.
HIGEN RNM’s Human-Friendly Actuator Platform: six joints from 60 to 348 Nm, sub-100 ms local force response, CES 2027.
A 1 September paper from IHMC and UWF shows Alex and Unitree H1-2 running runtime-editable loco-manipulation: 34-second push doors, 45-second ball sorts.
KAIST, KIST, and SeoulTech used machine learning to print a rubber-like DLP resin that stretches more than 6×, then built a pneumatic hand around it.
NVIDIA’s Isaac Teleop MANUS plugin reads haptic commands and drives five finger motors, closing the loop on sim teleop.
Ex-Meta FAIR researchers released Isaac 0.5: 36 billion parameters, 35 robot systems, open weights, and a scaling law that trades video for teleop.
Petoi launched Quaddle on Kickstarter Sept 2: a 4-servo omnidirectional quadruped kit from $99, running OpenCat.
Four university teams spent a month teaching ROBOTIS’s open K1 humanoid obstacle runs, K-pop, taekwondo, and parkour. Cohort 1 wrapped 31 August.
Skild AI’s S1 foundation model runs unseen tasks up to 10 minutes from a single video prompt, with no fine-tuning, the company and The Robot Report say.
Zhejiang University’s RSS 2026 gripper uses an active belt and two underactuated fingers to grasp books, paper, and film without finger gaiting.
Anthropic opened a research preview of MHS, a driver spec so AI agents can run robotic arms, liquid handlers, and other programmable gear.
DFKI and the University of Bremen’s B-Human won the RCAP Masters 9–0 and the WHRG humanoid soccer final 13–3, both against HTWK Leipzig.
CLAP learns action-conditioned video physics from mixed human and robot footage, then few-shot adapts to a Unitree G1 and a bimanual YAM.
Menlo Research says Asimov 1 now transfers flat-ground walking from simulation with no per-robot tuning, running 50 Hz on the motion board.
WIRED reports Meta is piloting Watney, Kinova Gen3, and ABB robots for cabling, power cycling, and reseating inside live AI data halls.
NSF is putting nearly $30 million into a UT Austin-led Science and Technology Center on how people and assistive robots adapt to each other.
NVIDIA’s Jetson Orin Nano 2 doubles entry-level edge inference to 78 TOPS and cuts 15 W power 40%. Module and kit are due first half 2027.
IEEE Spectrum’s Evan Ackerman talks with Persona AI about shipyard welding as a first job, with HD Hyundai and POSCO as partners.
Hugging Face’s Pollen Robotics opened $399 Microduck pre-orders: a 25 cm open-source-software biped that walks, skates, and trains in sim.
A closed-loop LQR controller learned from open-loop falls lifts unassisted walking on the standard 3D-printed Poppy humanoid.
A bimanual robot with three-fingered hands learns cascade, tennis, half-shower, shower, and box juggling in under five minutes on hardware.
A Science Robotics paper puts CTU Prague’s iCub in a crib with a mobile, using prediction and curiosity to model how babies learn their limbs matter.
Sharpa’s North humanoid opened at DQ’s Wujiang Road store on Aug 29, running a 55-step Blizzard on the shop’s existing gear.
A new perceptive locomotion stack completes a 1.5 km outdoor route and mixed indoor terrain using only chest depth and proprioception.
Tennibot opened Partner Lite orders at $1,495: a 25 lb stationary AI ball machine with V2’s app and drills. Intro pricing ends Sept 24.
Tohoku and VISTEC used inverse RL on a handful of stick-insect steps so VISTEC’s RedMirror hexapod learned walking, rough ground, and a missing leg.
ANTARA: Tsinghua’s Huoshen team won the WHRG large-robot 5v5 soccer final 10–1 over SPQR of Italy on the last day in Beijing.
HKEX filing: UBTECH’s H1 humanoid line did RMB 590 million on 921 units, +1,445%, while the group still lost RMB 339 million.
Olivier Laflamme disclosed two root RCE chains on Unitree’s G1 EDU, including a BLE path. The cloud key-recovery check is patched.
Xianglu’s 3K Vision cooking robot uses three 40MP cameras and CookingMuse to retune heat, time, and stir when the dish changes.
AgiBot’s X2 takes WHRG 100m obstacle gold in stock config; the company finishes 18 gold and 46 medals, edging Tiangong’s 45.
Berkeley and Stanford’s BeyondMimic, now in Science Robotics, tracks 2.5 hours of motion and deploys cartwheels and spin-kicks on Unitree hardware.
Gaoqing Power’s 80 cm, 13 kg ‘Little Milk Bottle’ jumps ~19 cm and enters 100 m, 400 m, obstacles, and martial arts.
X-Humanoid’s Tiangong Ultra wins the large-size 100m final in 8.64s, after 9.39 and 8.86 earlier in the week.
Team Tiangong / Tianzhuo wins the WHRG standing high jump at 3.4 m, judged on height, landing, and a tag on the lowest body point.
Beijing’s Tianzhuo team takes the WHRG standing long jump final at 4.83 m, a separate event from the 7.97 m running long jump.
Adelson’s lab posts a snap-on vision-based fingertip whose TPU skin survives hours of sanding and days of 39 N probes.
A Meta and Northwestern glove uses multi-core shape-sensing fibers to reconstruct hand pose at 4.9 mm fingertip error after one calibration.
Figure AI comes out of stealth with Index, a creator app that has taken 16M videos and paid $15M toward Helix training.
MiGA packs 103,000 demos across five gripper types. GVLA uses that to beat a π0.5 baseline by 7.62 points on mixed-gripper tasks.
Team Tianzhuo wins the WHRG 1,500-meter gold in 2 minutes 21.63 seconds, with the top three all under the human record.
Beijing Humanoid–HUST’s TianGong lifts 16 kg overhead at the World Humanoid Robot Games after a 15 kg start.
Spirit G2 took WHRG fire and library scenario golds. The extinguisher is 4.5–5 kg; AgiBot hung it from the wrist into a 20-degree bucket.
An IROS 2026 paper uses a world model to distill loco-manipulation into a vision student. On OMOMO it hits 92.2% success in simulation.
LD4WAM pretrains on 5,086 hours of human and robot video, then runs grippers and dexterous hands. Real-world average success is 70.5%.
A University of Tokyo IROS 2026 paper wraps Musashi-W’s elbow in layered silicone skin. Skin-only angle error is about 3 degrees.
A Beijing Innovation Center robot won WHRG long jump at 7.97 m and the 400 m obstacle in 4:11.44, a new test of whole-body control.
Hunan University’s TONAV pairs last-meter navigation with action-velocity chunks on a Unitree Go2-W. Manipulation success hits 80% with P–V teleop.
Beijing added eight dexterous-hand events. AgiBot said it reached every final and took seven golds, including beans with tweezers and block stacking.
Twelve teams played autonomous table tennis on the same 1.73 m humanoid. Organizers say the swing window is about 0.3 seconds.
Honda’s 16-DoF multi-fingered hand holds 50 N at a fingertip and is entering a RIKEN/NEDO Physical AI project aimed at harness work.
The Robotics and AI Institute’s Koala pair matches a handheld data gripper to a backdrivable robot hand for tool-use learning.
A PPO sit-to-stand policy on a simulated Unitree G1 hits over 97% success across eight chair heights, no motion capture required.
X-Humanoid’s Tiangong Ultra took the Large Group 400m final in 38.15s; Omni won the small group in 45.66s with a learned high-arm gait.
A visuo-tactile model trained on 1,800 human grasps of 60 objects hits 95% ID grasp success by predicting mass, friction, and stiffness.
XPeng’s robotics unit closed a $900M+ first round at over $6.3B, led by IDG, with Iron aimed at late-2026 production.
CAS Institute of Automation spinout CasiaHand shows six hands at WRC, including an M6 cell running a 24-hour pick livestream.
CEA-List’s CoToGrasp, headed to ECCV 2026, synthesizes Allegro grasps conditioned on contact topologies without object-specific training.
Beijing’s DaxAI showed the rideable Qiji X1 quadruped at WRC, a saddle-and-handlebars machine rated to carry 660 pounds.
Tsinghua, Shanghai AI Lab, and DEEP Robotics freeze a FastWAM backbone, cut action MSE 21.7%, and ship the ARMDOG dataset.
GAC’s Huilun shows fourth-gen wheel-legged GoMate Mini at WRC 2026, with ~50 units already on patrol in four cities.
Generalist’s GEN-1.5 hits 59% one-shot success across 10 short tasks with no gradient updates, then 83% after 10 steps.
Science Robotics publishes SONIC: 42M parameters, 100M+ mocap frames, 21k GPU hours, one policy for VR, video, and VLA whole-body control.
PartialBiGrasp generates dual-arm grasps from partial point clouds instead of assuming a full object scan.
Beijing’s RoboScience debuts the wheeled REX G1 at WRC: 22 DoF, 10 kg dual-arm load, 2,070 TFLOPS, hot-swap in five seconds.
UIUC’s self-demonstrated VLA method keeps instruction following on ALOHA: 90% cube pick success with 14 minutes of expert teleop.
Symbiosis Robotics’ Direct Perception Control demo has a Unitree-class humanoid climb in, steer, and throttle on a closed track.
X-Humanoid debuts a lightweight Omni at WRC 2026, plus Pelican-Unify, and parks TianGong 3.0 at a Mercedes-Benz stand.
X-Humanoid’s Tiangong Ultra clocks 9.39s in a WHRG heat, faster than Usain Bolt’s 9.58, with Honor’s Lightning at 9.47.
X Square Robot’s WALL-B model plus in-house 6-axis arms hit 1,816 parcels/hour at >98% in a livestream, then ran the cell at WRC.
A new paper trains humanoid tennis serve and rally styles from Nadal, Federer, and Djokovic video, then serves in the wild on a 1.7 m Atom.
Science Robotics cover work on a 1.2 m Booster T1 ties onboard vision to locomotion, with ~90% frontfield kicks and RoboCup use.
At WRC 2026 DexForce showed wheeled W1 Pros packing phones; Reuters says the same robots have run at a Lens Technology plant since early 2026.
CMU and MIT post Feedback Sampling MPC, a hybrid sampler that keeps MPPI stable on contact-rich humanoid walk and pick-and-place.
Galbot unveils ET1 at WRC 2026, a bipedal humanoid trained for tennis and scheduled for the World Humanoid Robot Games opening.
CEA-List’s IROS 2026 grasp planner trains only on gripper geometry, then hits 86.93% on MultiDex and 11 real YCB objects.
CAS and ImprintX train a hierarchical tactile world model that lifts chip, eraser, and USB success from 31% to 72% on a real arm.
HKU, BIT, and Tokyo model foot-by-foot granular contact and train a biped that identifies stiffness online, including 1.2 m/s foam trials.
A GelSight Mini plus a Fluid Reality display cuts teleop-to-human trajectory gap 29–79% when operators feel where contact actually is.
At WRC 2026 the 1.71 m L7 (55 DoF, 4 m/s) shares the floor with XHAND and a logistics stack ROBOTERA says already runs at SF and China Post.
Unitree lists a desktop 7-axis R1 arm at $1,650 / 9,900 yuan, 2 kg payload, 650 mm reach, with force feedback and open interfaces.
SJTU’s Video2DoorTraversal rebuilds a door from one RGB clip and hits 96.57% real-world traversal in about 13 seconds.
The second World Humanoid Robot Games draw 2,056 robots from 16 countries, with new tug-of-war, table tennis, and halved scores for teleop.
At WRC 2026 Xiaomi showed a 1.70 m, 66-DoF CyberOne after factory internships: 98% on nut loading, 90% on covers and crate folding.
NVIDIA and Michigan’s ADEPT pretrains dexterous RL, then zero-shots peg insert and dish-rack on Kuka-Allegro and Flexiv-Sharpa.
AVIC showed Fengxing-001 at WRC: 1.8 m, electro-hydraulic joints, 20 kg carry, and staff say it is in simulated aircraft-pilot tests.
TUM’s CompCPZ keeps the ‘or’ in robot commands, winning 1,900/1,918 sim trials and 12/12 Go2 runs where convex goals miss.
Tsinghua and GigaAI train a humanoid tracker as a world model. 81.3% terrain success, 99.3% fall recovery, G1-to-L01 transfer.
A Fujian–Osaka team conditions ACT on a metric grasp point. 88% success on a citrus mock-up; RGB-only ACT, Diffusion Policy, and SmolVLA grasp none.
USTC’s LabDex benchmark trains Franka + XHand on pour, insert, and stir. π0.5 hits 0.57 on atomic skills; no model finishes the full workflow.
LTA will trial Olly at Downtown Line Little India MRT, weekdays 10am–noon from Aug 31 to Sep 11, with staff on site.
China’s NORINCO debuts Fuxi at WRC 2026, a full-size teleoperated humanoid with an “external avatar” stack.
Pudu launches the MP2000 pallet robot: 2,000 kg payload, 20-second fork-in, 3D LiDAR-SLAM, and no site rebuild.
Berkeley’s Goldberg lab lets Claude Code solve Push-T with no demos: 100% success, 46% fewer steps, then 99.4% on 26 letter shapes.
A UCLA-linked team edits human manipulation clips into seven robot hands, then drives a real Franka from the decoded 3D states.
A 4,000-demo visuo-tactile suite scores pick-and-place only if deformation stays in a pre-calibrated window. Touch helps more under shift than in-distribution.
X-Humanoid’s Tien Kung 3.0 inspects high-voltage cabinets at WRC with State Grid’s model; LingLong L2.0 demos firefighting without a remote.
Vita Power’s ATOM debuts at WRC 2026 at 160 cm and 31 DoF, with fabric, a face screen, and on-site reservations.
HIT and Honor route paired stereo into a frozen VLA on a 33-DoF humanoid, hitting 60% full-task and 100% grasp on 100-second jobs.
UCLA trains visual loco-manipulation in 150k simulated scenes, then deploys zero-shot on a Unitree G1 at 73.3% success.
A System-2 service agent on Unitree G1 hits 92% on atomic skills and 63.3% on full tasks, and beats ReAct by up to 18.9 points in sim.
NVIDIA researchers condition a world model on camera-plane trajectories, cutting motion error and replaying RoboLab success at r=0.96.
Seoul National University’s λ-hold controller trains a 90-muscle model to sprint in about an hour using a stretch-reflex command, not raw excitations.
Northwestern’s ManiGuard-Bench finds 6–21% of successful VLA rollouts still violate a formal safety spec, then ships 8,000 annotated demos.
NineRay Technology launches the 173 cm RayNex G3 at Hong Kong Science Park, with 32 DoF and a 60 kg stage carry.
HiPHI is public on Hugging Face: 132 performers, 90 Hz optical MoCap, and G1 policies that run, crawl, carry, and pull a suitcase.
Peking University’s open PRISM dataset logs 45 hours of contact-rich factory skills across Franka, Realman, and LEJU platforms.
Agibot Finch’s hierarchical VLA searches subtasks with a world model, trained on 40,115 hours, and lifts long-horizon home-task success.
Tsinghua and NTU bolt a fast impedance reflex onto DP, π0, and DreamZero. Contact-stage success jumps 20–60 points without backbone fine-tuning.
AP: the Hangzhou humanoid maker raises 6.1 billion yuan, prices at 150.80 yuan, and is still up nearly 500% by midday.
University of Tokyo researchers train waypoint-guided RL on hook hands, then zero-shot a 19.6 kg upper body across four bars.
AP from the floor: Unitree boxes and plays table tennis, UBTECH prices a face at 168,000 yuan, and a helper robot still cannot fold a shirt.
A training-free recovery loop plans on synthesized cameras, then moves only the objects it has to. Real xArm6 success jumps as much as 85 points.
JSK’s IROS 2026 Detachable Wire Drive parks the motors in one base and lets arms and grippers click on through Vectran tendons.
Moxi 2.0 is shipping to named U.S. hospitals with a new world model, 9-hour runtime, and faster perception.
Peking University’s RA-L planner walks a pneumatic suction worm over 3D surfaces 12× faster by reusing short validated moves.
Iowa State researchers put a G1 and Dex3-1 on a football and report 93.6% spin efficiency with a follow-through controller.
HKUST-GZ’s underactuated hand uses six servos and a spring-stretch model to get sub-degree joint predictions without extra sensing.
Peking University and Beijing Innovation Center adapt frozen VLAs to whole-body loco-manipulation with a three-stage action flow.
TU Dresden and Stanford open-source a vision-based tactile probe that classified tissue at about 84% in an in-vivo trial.
UC San Diego trains a residual force tracker so an XHand can copy a human grasp without crushing a paper cup or missing contact.
SJTU and Shanghai AI Lab box in a 32-D latent space, then put the policy on two real Unitree G1s with red gloves.
HKUST researchers put 60-taxel pressure insoles on a Unitree G1 and walk foam and gravel the proprioception baseline cannot.
Tohoku’s ViHaTeleop is 0.7 kg of SLAM, a fisheye, and three LRA rings. Haptics raised success on every contact task they tried.
A residual layer on OpenVLA-OFT uses touch only to pick and scale a small correction. UniVTAC success jumps 30.6 points.
RAI Institute and Boston Dynamics train one motion-imitation recipe from MoCap, video, and animation, then drop it on Atlas, G1, and Spot.
A new VLA paper pairs a 100k-trajectory OmniShare dataset with a 15-joint action space for human-to-robot transfer.
Astribot’s ART fine-tunes π-FAST with LoRA so the policy can invoke vision and camera tools on 30k trajectories.
SNU and KTH flew a planar space-analog robot over 7,450 km of internet delay and still completed a docking run.
A Haar-split VLA ranks same-behavior action chunks so the robot does not jump when inference finishes mid-motion.
Jilin’s Isaac Sim benchmark covers 100+ flat objects and three grasp strategies, then transfers to a Baxter.
A RA-L ObjectNav stack orders household objects by LLM affinity, then scans them with Mamba. It hits 83% success in AI2-THOR.
Purdue’s hint² uses high- and low-level world models to satisfy LTL on CALVIN and a real UR5e.
A diffusion planner grounds image-space goals in local occupancy and lifts 5–8 m sim success from 21% to 72% over NavDP.
A new toolkit turns robot rollout videos into progress curves, then grades near-misses, recoveries, and messy successes.
SJTU’s 1B ReflexVLA hits 50.4% on a new six-task latency benchmark and 16/20 on a real conveyor.
Industry sources say RX ran an Aug. 10 prototype demo at Umyeon, including camera work. Public stills remain the 2022 Bot Handy.
V-Trion and ZHAW coat silver yarn in PDMS and show a 4×4 textile skin on a Franka arm with 403 ms reaction.
Ohio State, Georgia Tech, and Georgia State pair VR rehab games with a 270 g, 7-DoF robot coach for kids with CP.
Unitree’s official clip claims a 2 m standing jump and 12.66 m/s sprint from a prototype built in about three months.
Noida’s Addverb and EndureAir show a Sabal-20 airlift of the Trakr quadruped. Product specs are on Addverb’s site.
A KAIST RA-L paper makes continuous-time VIO spend spline knots on fast motion. Mean RMSE is 0.212 m on a UAV-style benchmark.
A new solver treats contact as complementarity and still uses a Riccati recursion. Closed-loop Push Box hits 50/50 with a 2.08 ms median solve.
An IROS 2026 paper generates Isaac Sim and Gazebo worlds from domain rules, including robot charging zones. The team made 450 layouts.
SJTU’s CARLA dataset has 6,933 outdoor routes with live traffic. Success drops more than 10 points versus older static street graphs.
Molex samples a hybrid connector that combines 15 A and 1 Gbps Ethernet for humanoid wrists, knees, and necks. Production is slated for late 2026.
IEEE Spectrum flags Monumental robots laying about 20,000 bricks on a house. The Amsterdam fleet already has walls on 100-plus sites.
Leipzig and Polish Academy researchers lift SAM 3 labels into a NeRF so a robot can query apples, branches, and free space in a real orchard.
Shenzhen startup Arkshel shows MX01 test footage: one platform switching among humanoid, quadruped, and aerial modes.
A George Mason UIST paper records gaze on two robot feeds, then helps the operator label why they looked away. Twelve people tried it.
CUHK researchers put strain gauges in a capstan drive so a 3.5 mm continuum robot can sense shape and contact without stuffing sensors in the drum.
AMAP’s action-conditioned video world model topped WorldArena 2.0 Track 1 by tying each arm’s SE(3) path to the generated frames.
ETH Zürich’s ECCV 2026 paper estimates dense contact and 3D force on hands and objects from one egocentric RGB image plus a mesh.
A new Habitat benchmark asks a robot to find an object by function, not name. FUSE picks extra viewpoints and hits 72% success versus 42% from a frozen camera.
NTU’s PassGen builds a 5,000-clip RGB-D handover set and an intention gate that cut false triggers from 25/30 to 2/30 on a real arm.
Karlsruhe researchers model end-of-life disassembly as a POMDP, then let arms unscrew, mill, or switch tools when a part is stuck.
CMU’s CoRL 2026 submission fine-tunes π0.5 on a single demonstration, then residual RL lifts success on a bimanual YAM in about six hours.
Cambridge and Utrecht’s IROS 2026 paper splits a scene into room vs people so a social robot can keep learning which actions fit each space.
Tsinghua’s plug-in rotates the end-effector’s relative velocity so a Flexiv arm misses a box instead of freezing the way classic potential fields do.
Czech startup REVEL shows Genesis, a special-purpose humanoid, plus Neural Gambit, a forearm sleeve that records expert grip and intent.
A zero-shot navigator builds room snapshots on the fly, then walks to a better viewpoint before trusting a match. It also ran on a Jueying Lite3.
Michigan’s ARM Lab built a compliant stem climber with PIN-SLAM and a 7–33 mm grip range. IEEE Spectrum featured it this week.
A coiled-spring pneumatic gripper pinches, rotates, hooks, and grasps without copying finger anatomy.
Vanderbilt and Sydney train a latent world model that vetoes a chunked policy at the last pre-contact step.
A 447-gram dual-quadrotor aerial manipulator puts out 1.019 N·m and tightens an M8 screw while holding pose.
SJTU and Shanghai AI Lab score 11 video models on whether a human demo becomes a usable robot tape.
VinMotion’s Isaac Sim benchmark walks G1 and H1 through 933 episodes instead of teleporting them.
Galbot, Tsinghua, and partners post a 153-hour mocap benchmark and HumanScore, a preference metric that beats MPJPE.
LG and NVIDIA sign an MOU for a bipedal humanoid on Isaac GR00T, Jetson Thor, and Halos, plus a CLOiD factory test in Tennessee.
Sydney and Vanderbilt train inner finger skills, then an outer policy that drops the copilot at deploy.
A NeurIPS 2026 competition from CUHK Shenzhen and DexForce trains on generated data, then scores policies on dual AgileX Pipers.
A hierarchical world model hits 98.7% on SurRoL PegTransfer by planning over variable-length events, not every primitive step.
KTH and Freiburg train a DINOv3 ROI readout with no gaze labels and lift real-robot success from 48% to 77%.
CAS researchers split VLA-RL credit by detectable task stage on RoboTwin 2.0 and LIBERO-Long.
Cornell shows leftover time in a trajectory is linearly decodable from PaliGemma activations, and useful as a stall detector.
The second World Humanoid Robot Games (Aug. 22–26) lists 666 teams and new events from tug-of-war to hotel housekeeping.
G0.5 is an autoregressive VLA that emits chain-of-thought and motor tokens together, beating π0.5 on R1 robots and BEHAVIOR.
A 200-million-edit benchmark asks image models to swap a human hand for a URDF-specified dexterous robot, then scores 11 editors.
Jung and Park measure a policy-induced hand prior across 17 initial arm poses and show more pose coverage in training cuts the bias.
A runtime recomposition stack on Spot, a Panther, and a Turtlebot swaps cameras and compute without a developer, including a reactor demo.
SMPC in simulation writes the expert data; sparse-reward RL then beats the teacher on a Spot with an arm and a Unitree G1.
StellarEdge’s VLA ranks first on VLA-Arena at 0.63 by retrieving a structured demonstration instead of fine-tuning.
Chongqing and Auckland researchers lift bins and balls with an H1-2 plus Inspire hands from a single RealSense view.
Seoul National’s skill codebook bridges what and how, then a 60 Hz policy adapts across action spaces.
Zhongke Huisi launched L1, D1, and M1 hands in Changsha, with an F-series tendon preview and a $2,700 L1 listing.
An IROS 2026 paper splits crowd-following into cost constraints and deploys it zero-shot on a ROSMASTER X3.
UC San Diego’s causal WAM uses failed rollouts as consequence labels and lifts real bimanual success from 82% to 89%.
UW and AI2’s 6B WAM beats π0.5 on real bimanual tasks and can drop to a 60 ms action-only mode.
A training-free logit-margin gate recovers OpenVLA accuracy on LIBERO-Long while keeping about 80% of the cache savings.
Koditschek’s lab recruits pitch-stabilizing hip torque to replace tail energizing, with hardware speeds of 1.02–1.77 m/s.
A three-stage whole-body planner plus residual RL threads a Unitree G1 through Cr < 1.5 openings where spline baselines fail.
The European startup left stealth Aug. 5 after a sub-90-day factory deploy on Unitree G1s, selling labor instead of robots.
A 16-person study on a LWR iiwa shows observer-based active support cuts effort and wins the agility test.
HKUST(GZ) and Midea put Semantic-3DGS on a Unitree quadruped and lift clutter success from 46% to 74%.
A Cosmos Policy WAM lifts SurRoL success from 63.5% to 77.8% after action-free video pretraining on a fixed label budget.
TermiTech’s RMC2 team loaded 25 shirts in ~23 seconds each at ICRA’s deformable-manipulation challenge.
ECCV 2026 framework hits 91% seen / 83% unseen-category success in sim and 90% / 81% on a real Psibot hand.
9-DOF three-finger DDD hand: ~18 N grasp, 4.7 N fingertip force, 50.1 N/m max passive impedance, patent pending.
Agentic harness lifts RoboMemArena full-task success to 59.6% by closing the prediction–deliberation gap.
Latent world-action model hits 79.2% on LIBERO-Plus without robot pretraining; π0.5+JEPA reaches 86.3%.
Soft-center, stiff-side anthropomimetic pads beat uniform soft pads on approach and rotation ranges; nail-free tips failed every trial.
MAB-RRT hits 100% on 76 Automate single-part extractions; PEEL dismantles 10–17 part assemblies on a Fetch simulator.
Online part-level reconstruction hits 83.4% mIoU and 21/24 task successes without CAD models.
RSS 2026 paper: 128 captures, 257k frames, contact-driven reconstruction, and closed-loop soft hand on a mannequin.
Alibaba DAMO open-sources a value foundation model that uses temporal distance, not preferences, and lifts real-world RL success.
reBot-DevArm ships 6+1 DoF arms with open hardware, ROS2, and LeRobot; DM kit lists around $1,799 assembled.
0.47B latent interaction model hits 97.5% on LIBERO and 77.45% zero-shot LIBERO-Plus with 60.6 ms H100 latency.
Offline teacher rollouts plus IDM pseudo-actions lift novel-task success on RoboTwin and real bimanual arms without test-time video generation.
IROS 2026 paper shows a battery multi-sensor platform and edge CV over 5G, with hand detection mAP 97.7% and RTT down to 12 ms.
Open ecosystem unifies 42 robot policies behind one schema, cutting integration from 5+ hours to ~30 minutes with agent skills.
Boston Dynamics says production Atlas runs ~4 hours typical duty and can autonomously swap batteries in under three minutes for multi-shift work.
AtlasVLA pairs a 4D voxel world memory with ego-working task memory, hitting 97.6% on LIBERO and +17.5pp on real long-horizon tasks wrist-only.
AutoIntervene watches action-chunking IL policies against a visual-action support memory and switches to the operator with calibrated thresholds.
C2Dex recovers stable object-side contacts from monocular video, retargets them to robot hands, and hits 57.78% success on DexYCB vs 17.78% for baselines.
A new arXiv method regularizes action-flow velocities across paired scene-camera views, lifting LIBERO-Plus camera-perturbation success to 87.2%.
CrossTracer scores 45.68 on NaviTrace—28% above Gemini-2.5-Pro—by refining VLA pixel traces for wheeled vs legged mobility.
Dyna Robotics says DYNA-2, trained on over a million hours of egocentric human video, hits 87% zero-shot customer quality vs 46% for Dyna-1.
IROS 2026 work from IIT uses asymmetric actor-critic RL plus learnable gait frequency so Kyon keeps moving when a joint loses torque.
IEEE Transactions on Haptics paper presents a multimodal fingertip with nail geometry for open-string, stopped-note, and harmonic tasks on the Chinese guqin.
HKU researchers propose kinematically coupled swept-volume planning so wheeled arms can haul large non-convex loads through clutter in real time.
New framework stores sparse 3D semantic memory so robots can chain multi-object searches across stairs—plus the HM3D-MFMON benchmark.
Mimic Robotics’ M1 packs 15 active DoF, >25 kg power grasp, and sub-0.05 Nm backdrive torque, with a matching U1 wearable for human demos.
HUMIT Lab's TECDAR uses a 2.5×3 mm 6D IMU at 7 kHz to detect and range extrinsic contacts on grasped objects within about 180 ms at ~7 mm accuracy.
TEMPO freezes the VLA backbone and runs two-timescale RL on a semantic projection layer and action expert, beating uniform RL post-training on CALVIN and real arms.
Politecnico di Milano couples preference-based optimization with modified DMPs so a UR5e rotates workpieces to match the painter’s hand — RA-L 2026.
A multi-exit video DiT planner scores trajectories early and skips full world-model depth, hitting 90.8 PDMS on NAVSIM at ~170 ms on an A100.
ATP couples RL musculoskeletal torque references with passivity-based cable-driven control; EMG shows up to 48% lower muscle activity vs no exo.
San Francisco’s Avatar Robotics closes a $6.5M seed led by AlleyCorp to scale semi-humanoid fleets, remote operators, and autonomy in live logistics.
CoMuDi plans multi-robot disassembly with ST-RRT* and temporal constraints, scaling to 49-piece assemblies and low idle times across six scenarios.
EvoHIL co-adapts success rewards, flow-matched action chunks, and relit offline replay so contact-rich policies keep working after illumination shifts.
GORDON turns action-free demos into graph-based object-centric rewards and auto-discovers subtasks, hitting 74.4% average success on long-horizon sim tasks.
A Bristol-led LfD study collects 3,142 touchscreen handwriting demos, extends GMM/GMR with force and time, and scores 71.5 human-likeness with open datasets.
PanoVLA fuses top-mounted panoramas with local cameras for wheeled dual-arm tasks, reporting 73.4% end-to-end success across four real-robot workflows.
ETH/ShanghaiTech Prior-SG builds hierarchical 3D scene graphs by aligning RGB-D instance maps with LLM prior graphs—no wall-based room heuristics required.
Skoltech’s SpaceVLA paints green/blue visual intent anchors on RGB so OpenVLA-7B hits 91.25% full-task success in closed-loop Unity pick-and-place.
Palo Alto’s Tacta Systems packages a 15-DoF fluidic-tendon hand, grain-of-sand tactile sensors, skill-capture glove, and Dexterous Intelligence AI into TactaBot.
A RO-MAN 2026 study maps finger motion to a Franka arm and cuts task time 53.5% vs joystick, with lower NASA-TLX, on remote surface-contact paths.
An input-ablation study of 32,874 model calls finds many VLMs score well while ignoring images; authors push selective hazard guardians, not monolithic control.
UIUC/Harvard ARGUS rebuilds multi-view scenes with VGGT, re-renders a fixed camera, and trains policies 4–6× faster on diverse viewpoints.
DLR’s ErgoSurf pairs online GPIS surface learning from touch with ergodic coverage so robots can sand, clean, or inspect without a pre-scan.
George Mason’s ICRA 2026 paper scores failure impact (probability × severity) and ships a MuJoCo injector for household drop and joint faults.
22 real-world task families, ~1,560 MoCap trials: Isaac Sim, Genesis, Newton, and six video models miss impact, cloth fling, and true accelerations.
arXiv:2608.04510 introduces GUARD, a test-time monitor that probes KV-cache grounding for diffusion VLAs and raises failure risk when evidence is thin.
Hierarchical Robotic Control freezes a Qwen planner, aligns an OpenVLA executor to subgoals, then runs hierarchical GRPO—98% on LIBERO-Long.
ASE 2026 paper IcFuzz fuzzes Isaac Sim with semantic stage segmentation and multi-level mutation, finding 11 bugs with 9 confirmed or fixed.
arXiv:2608.03701’s LiLa-WAM pairs compact latent foresight with Visual Transition Tokens, hitting 90.48% on 50 RoboTwin tasks under a single-GPU budget.
arXiv paper Mimir keeps scene belief and goal progress separate, then grounds them before each action—big gains on EmbodiedBench.
Beijing’s Noetix lists E1 at 136 cm / 40 kg with terrain walking, jumps, and stairs on the product page as terrain demos circulate.
ROBOTIS’s open AI Sapiens humanoid posts a Hangang riverside run demo under 38°C heat, stress-testing outdoor locomotion on Dynamixel-Q hardware.
arXiv:2608.05903’s Robust-WAM post-trains world-action models with DINOv3 future queries so lighting OOD hurts less without killing VAE pretraining.
SpikingNav’s SSE+SPN stack beats a matched ANN on ObjectNav and visual corruptions, and runs sensing on the Thruster-V2 neuromorphic chip.
Variable Impedance Diffusion Policy maps demonstration trajectory distributions to stiffness profiles for contact-rich manipulation.
DyPES-VLA learns cross-robot dynamics in a VLM and routes control through an embodiment-specific MoE head — 98% on LIBERO in the paper.
Tencent Robotics X and Tsinghua’s GeniWorld conditions video world models on URDF-rendered robot motion for OOD scenes and synthetic policy data.
Researchers show free-form chain-of-thought hurts low-level VLA control; in-context post-training plus tool use hits SOTA on sim and real arms.
JoyAI-RA 0.5 uses dual action alignment across 53K+ hours of ego video plus sim and robot data on AgiBot hardware.
KILVO runs kinematic-inertial-LiDAR-visual odometry in a hybrid ESIKF built for bipeds, with code and datasets on GitHub.
ShanghaiTech maps GelSight-style Poisson reconstruction onto an FPGA pipeline: fixed 0.211 ms latency, 28 ms protective reflex vs 170 ms host path.
YC-backed Nori Robotics lists the L3 bimanual wheeled home robot at $1,688 with a $200 deposit and San Francisco assembly.
NTU/PKU/BAAI ω-0 couples latent foresight with whole-body action latents on a 40+ hour household humanoid dataset.
A 5,670-trial study shows paper signs can steer VLM robot planners; simple defenses cut most of the risk.
Tsinghua’s SkillMemo segments demos with MoE experts, banks skill gating profiles, and lifts π0.5 to 98% average on LIBERO plus real UR5e gains.
KAIST’s TRACE learns proprioceptive odometry with foot-aware attention — less position drift on soft and outdoor terrain.
University of Tokyo’s IROS 2026 VLAff trains one VLM on 204K EgoAffordance episodes for visual heatmaps, grasps, and trajectories—68% real zero-shot.
World-to-Wrist VLA forecasts task-conditioned future wrist latents for fine contact work at 80+ Hz on LIBERO and real CoBoT Magic tasks.
XEWorld’s cross-embodiment testbed finds action-conditioned world models act like 2D pattern matchers—visual distance beats kinematics for transfer.
BridgeVLA++ posts SOTA on memory manipulation benchmarks while keeping few-shot real-robot generalization from the original BridgeVLA design.
DOBOT launches LUMO, a nearly 1.3-meter bipedal companion robot marketed for outdoor-to-indoor terrain and multi-role home/education use.
CTU Prague, CU Boulder, and TU/e show a conformal electrical impedance tomography tactile skin with 6 mm mean error on a curved prototype.
SiMDex treats human ego video selection as a recommender system, lifting dexterous VLA success from 47.7% to 61.1% with under 5% of a 32M pool.
SUSTech’s StableMimic blends tracking and post-fall recovery on Unitree G1 with 100/100 push recoveries and real hardware demos.
Vulcan Robotics’ Sourccey is a 1.03 m wheeled dual-arm platform with open hardware, Pi 5 compute, and an August–September 2026 ship path.
At Advancing AI 2026, AMD launched Ryzen AI Embedded X100, Kria AI SOMs, and an open robotics developer platform combining CPU, GPU, NPU, and FPGA.
CMP adapts general motion priors to the current task context without skill labels, improving humanoid RL performance across five control tasks.
Symbolic hand codes cut quantization error by ~75% vs Hand Labanotation and enable training-free finger edit, repair, and robot retargeting.
Shanghai’s DroidUp shows Moya, a biomimetic humanoid with heated silicone skin, pupil tracking, and claims of ~92% human-like walking accuracy.
New arXiv pipeline converts first-person human manipulation videos into multi-morphology robot data and improves VLA out-of-distribution generalization.
A new study maps why ISO-style power-removal safety breaks for balancing bipeds and instruments the gap on a Unitree G1 EDU cell with Siemens fail-safe gear.
LLP grows sparse parkour seeds into terrain-paired skills and deploys a single onboard depth policy on a 90 cm humanoid without runtime skill labels.
Pose Flow Matching for Humanoid Robots uses a frozen flow prior and Pose Geometry Score to improve dynamic motion tracking without ordered clips.
Framework synthesizes humanoid manipulation from a single egocentric RGB-D observation using generated video, keyframes, retargeting, and VLM-guided refinement.
IROS 2026 paper: printable AGSD patches crash OpenVLA; Structure-Aware Robust Fine-Tuning cuts failure and lifts real PiPER success under attack.
Wristband patterns for confirmations and exceptions cut workload and win preference over sensory haptics in bimanual teleoperation studies.
Shooting for Contact uses contact-implicit multiple shooting so RL policies transfer zero-shot to G1 crawling and a 180° jump-turn.
Training-time 3D tracker features lift VLA success on LIBERO-Plus, RoboTwin, and four physical bimanual tasks without needing the tracker at deploy.
IROS 2026: Unified Visuomotor Targets encode motor control plus visual scene transitions—no architecture change, no extra data—and improve bimanual policies.
BAAI's BifrostUMI uses Pico VR plus UMI grippers to teach Unitree G1 pick-place, bimanual produce, throws, under-table dumps, and walking coffee delivery.
Tsinghua Project Instinct unifies perceptive locomotion and general motion tracking so a G1 can vault, dive-roll, and kneel-climb on uneven terrain.
Foundation Robotics demos a forearm-motor, tendon-driven hand that cups and catches a baseball, aimed at its Phantom humanoid platform.
HALOMI uses UMI-style human demos, an active neck, and a manifold whole-body controller to hit ~85% average success on Unitree G1 tasks.
SJTU's HERO agent collects 664 autonomous episodes across four tasks, then distills L1/L2 experience into closed-loop L3 visuomotor policies.
HiWET's hierarchical RL hits 12.4 mm sim EE error on Unitree G1 and transfers zero-shot for real-world loco-manipulation under arm motion.
Horizon Robotics open-sources HoloMotion for whole-body humanoid control, with MoE inference around 300 FPS and zero-shot G1 demos.
Meta, NTU, and UW introduce HumanCLAW-Bench: 1,218 egocentric find-navigate-interact episodes. Best VLM sits on the target in only 16.8% of trials.
NVIDIA GEAR’s HumanoidMimicGen generates G1 loco-manipulation demos from few teleops, lifting real-world co-training success from 51% to 71%.
Tsinghua and Spirit.AI's HuMI collects backpack-portable demos without a robot present, claims 3× teleop throughput and ~70% success in unseen scenes.
Ant Group’s Robbyant released LingBot-VLA 2.0 trained on 60,000 hours across 20 robot configs, with open weights and real-robot pilots.
USC PSI Lab’s SIMPLE pairs MuJoCo physics with Isaac Sim rendering for 60 whole-body humanoid tasks and zero-shot sim-to-real transfer.
Northern California’s Threehalves pairs a humanoid upper body with a four-legged base, modular tool mounts, and fail-safe pneumatics for dangerous jobs.
Berkeley-led CLIFT turns GROD’s managed SFT API into a closed-loop flywheel on Unitree G1, reaching 96–100% success without opening model weights.
OpenDriveLab’s EgoHumanoid co-trains VLA policies on robot demos plus in-the-wild human video, reporting +51% generalization on real loco-manipulation tasks.
Florida’s IHMC took its all-electric Alex humanoid off the gantry for an off-site Maryland demo on gravel, slick surfaces, and uneven pavement for the Office of Naval Research.
Ex-DeepMind Applied Robotics founders launch Reimagine Robotics: factory robots taught by workers, not specialist programmers, already in plastics and hard-drive disassembly.
OpenDriveLab’s RoboNaldo curriculum RL hits 0.73 m free-kick error from 3 m on a real G1, with 13.1 m/s ball speed and outdoor onboard perception.
MIT-spawned Tutor Intelligence runs Data Factory 1—100 Sonny robots plus remote tutors—to train Ti0 VLA policies for industrial bimanual work.
At ICRA 2026, AGILINK showed contact-rich balloon twisting with OmniHand and introduced OmniHand 3 Ultra-M: 20 active DoF, full tactile coverage, direct drive.
Shanghai researchers’ FA-RDP adapts diffusion sampling rate for contact-rich manipulation—81.7% average success on box flip, switch, and button tasks.
Genesis AI unveiled Eno: a wheeled, headless general-purpose robot with human-like hands and the GENE foundation model, targeting deployments by end of 2026.
NCKU researchers turn Veo-generated human videos into Unitree G1 reference motions, then RL-track tasks like boxing and pick-and-place without real demos.
San Diego startup Tactus AI showed Elsie at ADLM 2026: a two-armed wheeled humanoid built to run existing lab instruments under a lab’s own SOPs.
Shenzhen’s X Square Robot is open-sourcing layers of its WALL stack—robot-free data, WALL-WM, and Wall-OSS—aimed at general-purpose manipulation robots.
Hangzhou BCI firm BrainCo showed EEG-driven robot control under 200 ms and a data-collection stack aimed at embodied AI training.
EPFL’s reversible multi-finger hand detaches from its arm, crawls spider-like, and grasps from either face — published in Nature Communications.
CEO Brett Adcock says Figure 03 can climb a ladder end-to-end. Company posts already detail perception-conditioned stair control on the same stack.
Stretch 4 ships now at about $30K with omnidirectional base, richer sensing, and in-home pilots for people with mobility impairments.
UK startup Humanoid reports 42–128% throughput gains and near-99% success on industrial tasks via on-robot reinforcement learning.
HTD predicts future tactile latents and forces, lifting success ~91% on five contact-rich humanoid tasks with open code and video.
Japan’s first airport humanoid demo: JAL Ground Service and GMO AIR plan a three-year Haneda pilot using two Unitree robots.
LimX’s COSA 0.5 technical report details V³-0 on the 43-DoF Oli humanoid: fast-slow VLA, onboard WBT balance, and real-robot RL with expert takeover.
New arXiv paper pairs control-barrier functions with onboard vision so a humanoid dodges throws using only head-camera depth—95% on hardware.
USC PSI Lab, NVIDIA, and WorldEngine open-source Ψ₀, a VLA for whole-body humanoid tasks trained on ~800h human video and ~30h robot data.
Shanghai lab open-sources UFO, an unsupervised RL stack for humanoids with TeCH, multi-robot bring-up, and kick-recovery demos on Unitree G1.
ROBOTIS documents AI Sapiens K1 — a 1355 mm, 35 kg, 23-DoF humanoid on Dynamixel-Q actuators with RL, imitation learning, and a planned full open stack.
TechCrunch visits Encord’s San Leandro warehouse, where Zander Labs EEG headsets and muscle sensors join egocentric video for physical AI training data.
RGB-only navigation policies trained in Gaussian-splat digital twins transfer zero-shot from Isaac Lab to real offices.
UNC and Stanford researchers open-source Handroid, a 0.33 m, 2.05 kg platform that reconfigures between dexterous hand and desktop humanoid.
Menlo Research’s Asimov 1 is a 1.2 m, 35 kg, 25+2 DoF educational humanoid with open design, kit pricing around $20k, and a $499 deposit.
Naver Labs Europe’s ARCBRAIN Flow lets non-developers compose robot paths and alerts on top of Naver’s ARC robot OS stack.
Korea’s NC AI and listed CMES Robotics signed an MOU to pair world models and robot foundation models with industrial 3D vision automation.
NEURA Robotics’ Qualcomm-powered 4NE1 Gen 3.5 humanoid collapsed during a Computex 2026 demo; Qualcomm says a communication glitch triggered designed safety.
ETH Zurich Soft Robotics Lab’s ORCA hand is a tendon-driven, open-source anthropomorphic hand with tactile sensors, sub-8-hour assembly, and long-run demos.
NYU’s Ruka-v2 open-sources a tendon-driven hand with a 2-DoF parallel wrist and finger abduction, under $2,000 materials, with faster teleop than Ruka v1.
UK hand specialist Shadow Robot is cutting access barriers with competitive pricing, rentals, shorter lead times, and packaged DEX-EE / Shadow Hand kits.
SKT hosted a Physical AI and robotics roundtable with Holiday Robotics and seven other startups to plan PoCs and robot foundation-model data.
South Korea’s XYZ is running its dual-arm DEUX robot at LoungeX: restocking, packing, and collecting on-site behavior data.
AGIBOT says its 15,000th unit — a G2 industrial embodied robot — left production in June 2026, after a rapid ramp from 1,000 to 15,000.
Chery-incubated Aimoga marks 2,000 cumulative overseas deliveries across 60+ countries, with JD.com humanoid pricing at ¥285,800.
BMW and Figure put Figure 03 on logistics sequencing in Hall 52 after Figure 02’s body-shop run on 30,000+ X3 vehicles.
DeepMind’s Gemini Robotics 2 stack adds whole-body humanoid control, multi-robot teamwork, and on-device adaptation — with ER 2 open in AI Studio.
Khosla-led round values General Intuition at $2.3B as the startup fine-tunes models from Medal gameplay labels onto real robots in minutes of data.
Ulsan partial strikes and union demands meet Hyundai’s 25,000 Atlas plan and a 2028 Metaplant America start — company says robots aren’t the current talks.
Open VLA GR00T 1.7 plus Isaac Teleop data collection now ship inside LeRobot — with LIBERO scores NVIDIA reports up to 96.5% average.
In Incheon, B-Human beat HTWK 4–0 in the first full-sized 11-vs-11 humanoid robot soccer game — both sides on Booster Robotics hardware.
Sanctuary reports 99.5%+ success at 2.54s cycle time on a Tier 1 wire-plugging task — and a strategy to run Physical AI on existing robots, not only humanoids.
CnEVPost, citing Wall Street Journal sources, says XPeng aims for more than 1,000 Iron units monthly by end of 2026 and store rollouts in 2027.
Alibaba-affiliate Ant Group led a ¥500M pre-Series A in Zeroth and has invested in 12 humanoid-related companies since early 2025.
DoorDash Labs earns Part 135 certification and launches in-house drones, plugging aerial delivery into the same network as Dot robots and human Dashers.
EngineAI confidentially filed for a Hong Kong listing after a $1.5B valuation and a Shenzhen plant claiming a T800 every 15 minutes.
Generalist says GEN-1 now trains across five-fingered hands, tools, and ~9,000 gripper variants on 500,000+ hours of real interaction data.
Hexagon Robotics says AEON is performing production work at BMW Plant Leipzig, with a path to production deployment by end of 2026.
Interlatent’s open-source robotics SDK puts VR teleoperation on a QUIC path with in-browser IK, a safety gate on the robot, and free hosted control.
NEURA Robotics and Qualcomm announced a strategic collaboration pairing Dragonwing IQ10 processors with NEURA’s cognitive robot stack.
German humanoid maker NEURA raised up to $1.4B from Tether, Nvidia, Amazon, Qualcomm, Bosch, and others to scale cognitive robots.
YC S26’s Nori Robotics pitches manufacturing-cost mobile bimanual robots to US labs after the FCC Covered List hit foreign advanced robots.
UBTECH unveiled the mass-produced UWORLD U1 Series in Shenzhen, claiming 13,361+ orders and pricing starting at 119,800 RMB.
China’s securities regulator approved Unitree’s Shanghai STAR Market IPO registration, a key step toward a ~¥4.2B public raise.
America’s first high-volume humanoid plant is live: 58,000 sq ft, 200+ staff, $20k buy-in or $499/month.
Tendon-driven, force-transparent hands: 22 finger/palm DoF plus a 3-DoF wrist, IP68, built for every NEO.
At WAIC 2026 in Shanghai, AGIBOT launched a full-size humanoid built for long shifts — plus industrial and dexterous systems aimed at real workflows.
Nearly 200 hires, $300M+ in multi-year Digit orders, and a Bay Area skill lab next door to Tesla country.
Digit's maker plans a SPAC merger with Churchill Capital XI, more than $300M in Digit v5 orders, and ticker AGLT if the deal closes.
Hangzhou startup AheadForm shows Origin-series heads built for blinks, eye contact, and speech-synced expressions.
Shenzhen’s AI² Robotics closed a massive financing round for AlphaBot wheeled manipulators and its Alpha Brain VLA stack.
Austin’s expanded Robot Park puts fleets of Apollo 2 to work collecting real-world data with Google DeepMind in the loop.
The Cloud Kitchens / Pronto stack rebranded as Atoms brings in Bain, Fifth Wall, and Uber — with Ben Horowitz joining the board.
China’s largest EV maker told China Securities Journal it will unveil its first humanoid robot in August at Di Space experience centers.
Shenzhen’s URKL puts full-size T800 robots in freestyle combat — 32 teams, live bouts, and a deliberate stress test for balance, sensing, and recovery.
The stealth-exit lab is putting 100+ AI robots online so the world can teach it how humans actually want to boss machines around.
July 28 rules block new foreign-produced advanced robots from U.S. equipment authorization—humanoids and quadrupeds included—alongside power inverters.
CEO Brett Adcock’s June chart: robot count climbing past human headcount as BotQ production ramps — a flashy milestone with real caveats.
West Egg Labs’ app books a humanoid for ~3 hours at $150 flat — no human cleaner in the room, remote help still possible.
Generative Bionics ships Gene.01 with full-body touch sensing and an open digital twin on pip, conda, and ROS.
Founded late 2023, the company says it has shipped more than 10,000 units — M1 and L1-series platforms built to leave the lab.
The CMU-founded startup bolts intelligence onto rented arms and skidders to install solar panels several times faster than manual crews.
Seoul’s Holiday Robotics closed a $105M Series A to commercialize FRIDAY, a wheeled industrial humanoid with hand-first design.
SoftBank exercised its put option. Hyundai Motor Group is buying the remaining stake and lining up Atlas for factory work by 2028.
Hugging Face’s LeRobot community released a full-stack, 3D-printed humanoid for robot learning — hardware, sim, and training included.
Shenzhen humanoid startup LimX closed a $200M pre-IPO round and is preparing a likely Hong Kong listing amid China’s IPO race.
Japan's Mitsubishi Motors and University of Tokyo spinout Highlanders aim to mass-produce bipedal robots at Kyoto by end-2027.
Ex-Optimus lead Jay Li’s startup settles with Tesla, raises from First Round and YC, and starts shipping a 22-DoF hand plus sensor gloves.
Sunday Robotics says ACT-2 folded laundry at 99.1% success across 785 zero-shot attempts in homes Memo had never seen — beta still fall 2026.
Invite-only house, deep, and office cleans: one hour flat, human-plus-AI control, bathrooms off-limits unless you opt in.
Business Insider’s July rundown: Gen-3 production “soon” in Fremont, supply-chain drag, and photos of the first Optimus line.
London-based Humanoid raised $152 million led by Prime Movers Lab, with Bosch lined up as contract manufacturer and industrial partners on the cap table.
Industry City neighbors Highline Commerce and Ultra Robotics put stationary dual-arm OP1 robots on a live fulfillment floor outside Amazon.
TRI spinout led by MIT's Russ Tedrake raises at a $1.1B valuation, putting wheeled humanoids on real factory shifts instead of demo stages.
Y Combinator-backed Weave Robotics prices Isaac 1 at $7,999 or $449/month, targeting California deliveries this fall and 1,000 U.S. homes by end-2027.
After buying SceniX, Fei-Fei Li’s World Labs demos real-to-sim-to-real policies that run an hour on hardware.