SayCan
Multiplies LLM semantic task-step probabilities by RL value-function affordance scores to select physically feasible robot action plans.
The loop
- Proposes
A language model scores candidate atomic robot skills by their semantic probability of advancing a high-level user instruction.
- Judged by
Pre-trained reinforcement-learning value functions (affordance critics) evaluate the exact probability that each skill will succeed from the current camera state.
- Steers
Each step takes the highest-scoring skill by the joint product P(say) x V(can) until the termination token is reached; there is no beam or tree search.
Headline result
Executed 101 real-world mobile manipulation tasks in an office kitchen with 84% planning rate and 74% physical execution success.
What this does not prove
Bounded by the repertoire of pre-trained low-level manipulation skills and cannot recover if an early step irreversibly alters the scene without replanning.
How it generalizes the harness
Bellman value-function affordance critics gate every proposed language-model step according to physical execution feasibility.
Classification
| What it proposes | Mathematical construction |
|---|---|
| Generative model | PaLM 540B semantic skill proposer |
| Algorithmic authority | RL Bellman value-function affordance critics and physical robot controller |
| Feedback | Score |
| What gets updated | Search tree |
| When it runs | At inference time |
| Validity | Empirical (measured) |
| Evidence | Experiments reported |
| Lineage | Original system |