A plausible plan can still ask a robot to do something physically impossible. The central insight is to separate knowing which action would help from estimating whether that action can succeed, then combine those judgments before acting.
Key Findings
- Language-model scores assess task relevance; skill value functions assess feasibility in the current state. Their combination grounds each choice in the robot's capabilities.
- On 101 instructions in two kitchen environments, the updated PaLM-SayCan system reports 84% planning success and 74% execution success. The gap shows that correct decomposition does not guarantee reliable control.
- Repeated skill selection supports extended instructions without asking the language model to generate motor commands directly.
Who Should Read It
Useful for researchers connecting high-level reasoning to a trained skill library. The approach assumes suitable skills and value functions already exist; it does not learn arbitrary physical actions from language. Feedback through current value estimates also leaves failure recovery less explicit than in later closed-loop planners.