NUS's APPL uses structural priors to bridge skill learning and composition in robot manipulation
NationalUniversityofSingapore · hf · 2026-10-02
Researchers at the National University of Singapore propose APPL (Agent Priors-guided Policy Learning), tackling how few-demonstration robot learning needs both compositional and skill generalization—two capabilities that depend on each other yet lose information between the composer and the skills it calls.
Core idea: each policy's structural prior (e.g., "grasping depends only on gripper pose relative to the object") becomes part of the interface—shaping where the policy generalizes during training and telling the composer where it applies in language.
How it works:
- A construction agent segments demonstrations into reusable skills, proposes multiple priors per skill, and trains/verifies one policy per prior.
- A runtime agent selects among prior-specific policies and composes them toward new goals.
On MetaWorld and long-horizon ManiSkill tasks, APPL improves out-of-distribution skill generalization and enables previously unseen skill compositions; ablating the interface information substantially reduces performance.
More from Embodied
- Boston Dynamics' new Atlas hand GR3 drops the pinky: 13-DoF design trade-offs explained — xiaohu · 2026-10-02
- Hugging Face LeRobot adds humanoid support: π0.5 policy drives 29-DoF Unitree G1 — freelerobot · 2026-10-02
- Researcher's blunt verdict on humanoids: "The robots do not work" — andrew_n_carr · 2026-10-02
- Robot's Early Taichi Policy Trained on Kaggle Shows Promising Kung Fu Moves — freelerobot · 2026-10-02
- BeyondSCe grasps objects by past-event reference, hitting 77% success on occluded targets — SeoulNationalUniv · 2026-10-02
- Open-source Harness fork moves coding agents out of the app into orca — dee_hw · 2026-10-02