A robotics preprint reports that a wheel-legged mobile manipulator opened and traversed five real doors after a smartphone RGB video of each target was used to build a simulated door twin. The reported door-by-door results were 35 successes out of 35 trials on two doors, 32 out of 35 on another, 33 out of 35 on a fourth and 34 out of 35 on the fifth, for an average success rate of 96.57%.
On three structurally similar unseen doors, the system recorded 25 successes out of 35 trials, 31 out of 35 and 29 out of 35, producing an average zero-shot success rate of 80.95%. The paper reports that this transfer required no additional training or trajectory generation.
From a phone video to a robot-ready door
Video2DoorTraversal is built around DoorTwin, which combines video geometry grounding, articulation-aware program generation, reference-view rendering and critique, and appearance transfer to reconstruct the door in simulation.
That recovered articulation feeds a simulation-in-the-loop agent. The agent converts it into a parameterized traversal program, diagnoses failed rollouts in the simulator, refines the program and produces executable demonstrations without human teleoperation.
Those demonstrations are used with ArticuACT, a policy that learns whole-body traversal from depth observations. It adds robot-centric Plücker conditioning—a geometric description of the robot-door relationship—and auxiliary supervision about the interaction state. At deployment, it uses only onboard depth and the robot’s own state readings, without privileged information.
How the tests were run
The hardware setup used a Unitree A2-W wheel-legged base with a Unitree Z1 arm and two onboard Intel RealSense D435 cameras. The study collected 200 successful demonstrations per door instance, and learning-based policies ran at 25 Hz.
In simulation, the methods were evaluated on 20 door instances over 256 trials, with independently sampled random seeds and the same randomization ranges used for training and evaluation. Door-opening success required reaching a predefined unlocking threshold and opening the door beyond 80° within 20 seconds. Full traversal additionally required moving at least 1 m beyond the doorway; a timeout or collision before unlocking counted as failure.
High scores, with weaker transfer to unseen doors
In simulation, the proposed method reported 98.44% door-opening success and 97.27% traversal success.
On the asset-generation benchmark, DoorTwin reported a Scale Score of 94.95, mIoU of 0.972, PSNR of 18.53, SSIM of 0.692, LPIPS of 0.408 and VLM Score of 56.74. Among the listed methods, it had the highest Scale Score, mIoU, PSNR and VLM Score and the lowest LPIPS.
On physical trials, the full opening-and-traversal sequence took approximately 13 seconds on average. In policy ablations, the complete model was reported as 17.58% higher than vanilla ACT with end-effector commands, 26.18% higher with joint commands and 2.35% higher than the strongest single-module variant under joint-command control.
A separate trajectory-generation ablation reported 85.63% success for the full agentic method, with 2.5 average loops. The rule-based, no-simulation-rollout and VLM-fix variants recorded 74.38%, 48.13% and 85.00% success, with 0, 0 and 2.8 average loops, respectively.
Traversal success in the paper’s data-scaling analysis rose from 59.38% with 50 trajectories to 97.27% with 200 trajectories.
What the evidence leaves open
The work is an arXiv preprint, version 1, dated 20 Aug 2026. No confidence intervals or other uncertainty estimates are reported for the simulation, real-door or zero-shot success rates.
The physical evidence is limited to the listed five-door evaluation and three structurally similar unseen doors on the reported Unitree setup. The study therefore leaves open how the approach would perform on other robot platforms, pull doors, additional handle mechanisms and more diverse door geometries.
Within those boundaries, the reported zero-shot result describes transfer to structurally similar doors; it does not establish broad cross-structure generalization.
Paper data and sources
Original title: Video2DoorTraversal: Push Door Traversal via Simulated Door Twins
Authors: Xincheng Tang, Yiji Chen, Youhan Xie et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text