Robust Autonomous Humanoid Loco-Manipulation

Develop robust autonomous loco-manipulation skills for humanoid robots.

Background

The paper addresses the challenge of enabling humanoid robots to perform complex, long-horizon loco-manipulation tasks autonomously and robustly. While reinforcement learning has shown success for legged locomotion, transferring these successes to manipulation-rich scenarios remains difficult due to planning complexity and environment interactions.

DreamControl-v2 proposes training a guided diffusion model directly in the robot’s motion space, aggregating diverse human and robot datasets to produce higher-quality reference trajectories for downstream RL. Despite these advances, the authors explicitly frame the broader goal—robust autonomous loco-manipulation for humanoids—as an open problem motivating their work.

References

Developing robust autonomous loco-manipulation skills for humanoids remains an open problem in robotics.

— DreamControl-v2: Simpler and Scalable Autonomous Humanoid Skills via Trainable Guided Diffusion Priors  (2604.00202 - Harithas et al., 31 Mar 2026) in Abstract

Throughout this work, we have wanted to add force-based action primitives. The codebase even includes a ``wrench action'' node, which was used to pick up boxes. Force-based action definitions could provide proprioceptive capabilities such as detecting when the door handle is all the way turned. It could also enable compliance when manipulating articulated objects. However, we still have not had the time to approach it properly from a controls perspective.

— A System for Fast, Resilient, and Adaptable Loco-Manipulation Behaviors on Humanoid Robots  (2609.01518 - Calvert et al., 1 Sep 2026) in Section “Natural Next Steps”

Cross-workspace stacking remains open, 0.5 stays low on pick-box, and multi-object transport is hard for imitation learning, demo-driven reinforcement learning and coding agents.

— BiGym 2.0: Benchmarking Learned and Agent-Developed Policies for Humanoid Household Manipulation  (2610.07594 - Zhang et al., 6 Oct 2026) in Abstract; Sections 1 and 4.4

We did not test whether agents improve by accumulating verified programs and reusing them across tasks.

— BiGym 2.0: Benchmarking Learned and Agent-Developed Policies for Humanoid Household Manipulation  (2610.07594 - Zhang et al., 6 Oct 2026) in Section 5, Discussion and limitations

Our evaluation covers a limited range of task horizons, and performance on longer-horizon loco-manipulation tasks remains to be established.

— Being-M0.7: A Latent World-Action Model for Humanoid Robots  (2610.11283 - Yue et al., 8 Oct 2026) in Conclusion, paragraph “Limitations and Future Work”

Human supervision also does not improve every behavior under the current training schedule: human-then-mixed training improves locomotion, but robot-only training achieves higher manipulation success. How to retain the locomotion benefit while improving object interaction remains an open question.

— VioLA: Learning Generalist Humanoid Control Policies from Human Data  (2610.12435 - Albaba et al., 8 Oct 2026) in Limitations and Conclusion, paragraph 'Limitations'