Compositional Humanoid Manipulation through Isomorphic Teleoperation and Independent Skill Routing
A Unitree G1 humanoid combined left-arm and right-arm skills it had only ever been shown separately, succeeding in 73.3% of six unseen pairings, while pi0.5, GR00T N1.7 and MoDE completed none, according to this preprint from Shanghai Jiao Tong University and partners. A low-cost exoskeleton for teaching the robot raised a full cabinet workflow from 35% to 65% success.
Can a humanoid reuse arm skills demonstrated one arm at a time in new left-right combinations, without anyone demonstrating every combination?
Demonstrating every pairing of two arms' skills grows combinatorially. Policies that predict both arms jointly entangle the skills, so a pairing absent from the data often fails even when each skill works alone. Contact tasks that use forearms and torso are also hard to teach with VR controllers.
Generalist policies such as Physical Intelligence's pi0.5 and NVIDIA's GR00T N1.7, and mixture-of-experts diffusion policies such as MoDE, learn joint dual-arm actions. VR teleoperation with inverse kinematics (VR-IK) maps hand-controller poses to joints, which does not control elbow placement or arm-torso contact directly.
HS-MoDP gives each arm its own semantic skill condition and router into a shared pool of diffusion experts, so each arm's skill stays selectable while parameters are shared. A vision-language scheduler alternates manipulation with repositioning by an existing walking controller. G1Exo, a wearable exoskeleton matching the G1's 14 arm joints with a chest plate as the torso boundary, records joint-level demonstrations. Each single-arm skill had 50 demonstrations; inference runs on one RTX 5090 GPU.
On six unseen tabletop and vegetable pairings HS-MoDP averaged 73.3% (60% to 85% per pairing, 20 trials each), and only HS-MoDP completed any pairing. Bimanual basket lifting from separately taught arms reached 65%. Training on G1Exo rather than VR-IK data lifted HS-MoDP's handleless-basket cradle from 15% to 85%, bottle gathering from 65% to 100%, and the full cabinet workflow (retrieve, reposition, retrieve while closing the door) from 35% to 65%. In a 10-operator study, G1Exo gave higher teleoperation success and shorter times than VR-IK on both contact tasks.
The authors say coordination under a shared load remains weaker than either arm alone and that concurrent execution can disrupt goals. We add that tests use one robot model, 20 trials per condition and a fixed set of household-style tasks; the scheduler depends on a vision-language planner whose errors are not separated out. It is a preprint, and the code and hardware are promised only after review.
The finding matters to Unitree G1 users and humanoid integrators that collect data by teleoperation: skill reuse cuts demonstration cost, and a cheap isomorphic exoskeleton beats VR controllers for contact work. Teleoperation-data vendors and companies such as co-author affiliation Xense Robotics are the near-term beneficiaries; factory use needs more robust shared-load coordination.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.