Learning General-Purpose Robot Policies

Develop methods for learning general-purpose robot policies.

Background

The paper surveys prior approaches spanning small expert models, robotics foundation models, and neuro-symbolic hybrids, and argues that none fully bridge high-level task reasoning with low-level continuous control across diverse tasks. This gap underlies the difficulty of creating broadly applicable robot policies that work without extensive in-domain data collection or fine-tuning.

The proposed Language Movement Primitives (LMPs) framework aims to partially address this by grounding vision-LLM reasoning in Dynamic Movement Primitive parameterization, enabling zero-shot manipulation through interpretable control parameters. Despite these advances, the broader objective of truly general-purpose robot policies is identified as an open challenge.

References

Learning general-purpose robot policies remains an open challenge.

— Language Movement Primitives: Grounding Language Models in Robot Motion  (2602.02839 - Dai et al., 2 Feb 2026) in Section 2 (Related Works), opening paragraph

Furthermore, zero-shot or few-shot transfer of learned physical actions such as different grasps and trajectories to new objects, environments, and embodiments remains an open problem.

— Toward Unified Robot Learning: Bridging Representation, Vision-Language-Action, and World Models  (2609.03927 - Mehta et al., 3 Sep 2026) in Section 3.2, Out-of-Distribution Generalization

RQ2: What functions must not be available beforehand, and how must the agent use context from the scene to achieve its goals?

— Deploying Foundation Models for Embodied Navigation  (2609.25666 - Dorbala et al., 22 Sep 2026) in Section “Conclusion {content} Future Directions,” subsection “On Day-One Unboxing”