HALO Teaches Humanoids to Coordinate With Human Partners During Shared Carrying Tasks

Carrying something with another person requires constant adjustment: slowing down, changing direction and keeping the load level when one partner stops or crouches. The HALO research project tackles that coordination problem with humanoid robots, demonstrating shared pushing and carrying tasks with a human partner.

Key Takeaways

Show all

• HALO explores how humanoids can coordinate with people during physically coupled tasks, including transport through confined spaces.
• Its mathematical stability result concerns how policies learn, rather than a blanket guarantee of safe physical behavior.
• Real-robot results are promising but limited: the experiments used motion capture and a small number of trials.

The work comes from researchers at the University of Texas at Arlington and Carnegie Mellon University. The project page identifies it as an ICML 2026 oral presentation. Its arXiv record dates the first submission to March 4, with a revision on May 31.

Learning to cooperate

The central difficulty is that independently learning agents can improve their own strategies while undermining coordination. HALO adjusts those learning updates to reduce disagreement with the shared objective. The authors distinguish this learning-stability result from methods that constrain physical trajectories for safety.

The project’s demonstrations cover pushing with precise orientation, moving through narrow spaces and handling long objects. Additional clips examine coordination when a partner is obstructed or changes height. Those are useful tests because a shared object makes each participant’s movements immediately consequential for the other.

What the physical tests show

According to the paper, training took place in Isaac Lab, followed by experiments pairing a Unitree G1 with a person using motion-capture support.

Results averaged over five trials show confined-space transport taking 76.2 seconds with HALO, compared with 81.5 seconds for PCGrad and 101.6 seconds for a scripted-partner baseline. HALO and PCGrad both recorded no drops in that condition. In the long-object halting test, both still had a 20% drop rate; HALO reduced post-halt drift relative to PCGrad.

These are encouraging comparative results, not evidence of failure-free collaboration. The limited trials and instrumented setup leave broader reliability unestablished.

The authors link a shared SafeHRC code repository, which covers HALO and a separate project called SAGE. Its detailed training instructions currently describe the SAGE configuration, a distinction worth retaining when assessing reproducibility.

For humanoid deployment, the practical question is compelling: can a robot become a useful physical teammate when the person beside it behaves unpredictably? HALO offers a research approach to that problem, with demonstrations and measured results that make the remaining gaps visible.