Add SayCan-style affordance grounding (embodied_ai/39)

[rsasaki0109]

What

A new example, examples/embodied_ai/39_saycan_affordance_grounding.py: a tiny, faithful SayCan loop (Ahn et al., 2022, Do As I Can, Not As I Say). The repo had no foundation-model loop; this adds the smallest honest one and ties it to the existing clarifying-question / conformal-ask-for-help line.

The lesson

A language model is a good planner and a bad robot. Asked to "wipe the table" it proposes pick up the sponge — the right idea — without knowing whether the robot is anywhere near the sponge. SayCan scores every skill twice and multiplies:

score(skill) = p_LLM(skill furthers the instruction) × p_affordance(skill works now)

The product is high only for a skill that is both useful and executable, so the greedy argmax walks out a feasible plan with no separate planner and never commands a skill whose preconditions are unmet.

The contrast is built into the same file via a ground flag (mirroring MCL's augment):

grounded (SayCan): go_to_sponge → pick_sponge → go_to_table → wipe_table (goal in 4–5 steps)
ungrounded (Say only): argmax LLM = "pick the sponge" from the wrong place, forever → timeout

The failure log tells the story directly: ungrounded racks up affordance_violation every step then timeout; grounded has none (or a retried skill_slip).

Contents

• A self-contained KitchenWorld (two locations, five stochastic skills with preconditions + affordances) and a SayCanAgent.
• The "LLM" is a small, transparent scorer conditioned on the running facts (held / clean) — the history-conditioned query SayCan makes — but deliberately blind to physical preconditions, which is exactly what the affordance term grounds. Comments make the stand-in explicit.
• Three smoke tests: grounded walks the feasible plan with no affordance violations; grounded retries a stochastic skill_slip and still wins; ungrounded loops on affordance_violation and times out while grounded succeeds on the same seed.
• examples/README.md row + a full examples/embodied_ai/README.md section; example count 41→42 and test count 115→118 in README.md / docs/status.md.

Verification

• Seeds 0–7: grounded cleans every seed in 4–5 steps (retrying slips); ungrounded never cleans.
• Full suite green (131 passed).

References

• M. Ahn et al., "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances," CoRL 2022 (arXiv:2204.01691, https://say-can.github.io/).

🤖 Generated with Claude Code