Robotics

One Phone Video, Then a Wheel-Legged Robot Walks Through the Door

Robb Harlan 4 min read

A Shanghai Jiao Tong University team posted Video2DoorTraversal on 20 August 2026: film a real door with a phone, rebuild it as a simulation twin, then send a wheel-legged manipulator through that door on its own.

The hardware is a Unitree A2-W base with a Unitree Z1 arm and two RealSense D435 cameras (head and wrist). Perception and policy run onboard, including a Jetson Orin NX for vision and arm commands.

Wheel-legged robot approaching, opening, and walking through real doors
Real-world approach, handle, push, and pass-through on several doors. Source: Video2DoorTraversal paper.

DoorTwin, then ArticuACT

DoorTwin turns one RGB video into a metric, articulated, simulation-ready door: panel, handle, hinge side, textures. A simulation-in-the-loop agent then writes a skill program (approach, grasp, rotate handle, push, pass, retract) and patches failed rollouts until the sequence is physically executable.

ArticuACT, built on action-chunking, predicts coordinated base velocity, yaw, six arm joints, and gripper from dual depth images plus a 99-D robot state. Extra tricks are robot-centric Plücker rays so pixels sit in the same frame as the wheels, and auxiliary predictions of contact, handle rotation, and door angle. Those extras are training-only. They do not change the control interface.

The numbers that matter

On five real doors, 35 trials each, the policy finished 169 of 175 runs: 96.57%. Per door: 35/35, 32/35, 35/35, 33/35, 34/35. Vanilla ACT on the same stack sat at 65.71%. Open-loop replay was 18.29%.

Average time for approach, open, and walk-through is about 13 seconds. Zero-shot transfer to three structurally similar unseen doors, with no extra data generation, was 80.95% (25/35, 31/35, 29/35).

In simulation, full traversal hit 97.27% against 64.84% for vanilla ACT. DoorTwin also beat Articraft, Articulate-Anything, and PhysX-Omni on metric scale and mask overlap for 20 real door instances.

Wrist camera frames of handle contact with plots of velocity and interaction state
Handle contact, then door angle, then forward speed as the robot commits to the doorway. Source: Video2DoorTraversal paper.

The authors are explicit about the hole: this is push doors. Pull doors, other handle types, and weirder geometry are listed as future work.

A Human’s Take

Filming the actual door instead of hoping a generic asset is close enough is the part I would steal. Thirteen seconds through a doorway is a useful clip. I will care more when the same stack pulls the door toward itself and does not eat the frame.

Sources