Robotics

Two Dexterous Hands, One Depth Frame: Real Bimanual Lifts Hit 73%

Shar Hendrix 4 min read

Large objects do not fit in one hand. Most papers that admit that stay in simulation and ask for a full mesh. A Chongqing University / University of Auckland team (arXiv:2608.10383, August 11) runs the job on a real Unitree H1-2 with two Inspire RH56DFTP hands and a head-mounted RealSense D435i. From a single segmented depth view they report 73.33% lift success on held-out objects.

Dual Inspire hands enclosing a pale plastic bin from both sides
Cooperative enclosure of a bin too wide for one hand. Source: Li et al., arXiv:2608.10383.

Teleop data, then a diffusion pose

They collected 353 grasps on 33 large objects (6–20 poses each) with Apple Vision Pro teleop at 15 Hz: commanded and measured joints, RGB-D, finger torques, and palm/finger tactile. SAM masks the first frame. A human marks the “just grasped, not yet lifted” frame. Code and data: github.com/zhangdana483/real_bi_dex_grasp.

A PointNet++ turns the partial cloud into a 1024-D code. A DDPM denoises a 26-D joint vector (two 7-DoF arms + two 6-DoF hands) over 200 steps. Joint limits are clipped. Execution is not a single snap: lift the arms, approach from a pre-grasp offset, then squeeze the palms together until they stall and curl any finger that is not yet in contact (torque >2.9 N or tactile above threshold).

What beat the baselines

Object-level 80/20 split. Each test object: 5 poses × 3 attempts. Success means a 5 cm lift, 2 s hold, no visible slip, inside 30 s.

  • Theirs: 73.33%
  • BimanGrasp-DDPM (sim poses, IK’d onto the real hands): 40%
  • ViSiL-HD (cuboid RL policy): 33.33%
  • GraspNet (two-finger emulation): 73.33% (wins on thin-rim buckets)
  • DexGraspAnything: 46.67%

Ablations: no arm squeeze 46.67%, no finger adjust 53.33%, no motion planning 20%, handcrafted open-palm baseline 26.67%. Single-hand methods failed on boxes and basketballs that have no pinch rim.

Overhead photo of 33 large objects next to the H1-2 humanoid and an Inspire hand
Dataset objects and the H1-2 plus Inspire RH56DFTP platform. Source: Li et al., arXiv:2608.10383.

Failures they show: table collisions if you skip the lateral lift, loose contact if you skip force refine, torso blocking a far-side arm, and a heavy box grasped off-center so the far end slips.

A Human’s Take

73% on unseen bins from one head camera is the kind of receipt I want from “bimanual dexterity.” The force close is doing a lot of the work, and that is fine. I would not ship this next to a person yet: timeout is 30 seconds, the scene is a clean table, and SAM plus a center-merge hack is the perception stack. Still, they lifted real balls and boxes on metal, not in Isaac.

Sources