PACTS: Joint Action-Predicate Modeling Enables Zero-Shot Skill Composition in Robotics
A novel method presented in arXiv (2605.20648) unveils Predicate Action Skills (PACTS), a set of closed-loop visuomotor policies that integrate action trajectories with symbolic predicate results. Unlike current generative policies that focus solely on action distributions, PACTS facilitates zero-shot combination of established skills without the need for retraining. By producing consistent action-outcome sequences within a unified model, PACTS enhances both action generation and predicate classification. This research tackles a significant challenge in Learning from Demonstration (LfD), where robots struggle to adapt to new skill combinations. The authors suggest that actions and their symbolic impacts should be modeled together, leading to more effective skill composition. PACTS develops internal representations that boost performance across both tasks.
Key facts
- arXiv paper 2605.20648 introduces Predicate Action Skills (PACTS)
- PACTS jointly models action trajectories and symbolic predicate outcomes
- Enables zero-shot composition of known skills without retraining
- Addresses failure of existing generative policies to generalize to new skill compositions
- PACTS is a class of closed-loop visuomotor policies
- Produces coherent action-outcome rollouts within a single model
- Joint modeling improves both action generation and predicate classification
- Work is in the domain of Learning from Demonstration (LfD)
Entities
Institutions
- arXiv