Physical AI Action-Outcome Experiment 013

Date: 2026-06-23 Status: Research complete Classification: Research-only

Goal

Compare similar robot incident retrieval, runbook/action-prior recommendation, reflection-only memory, and context-gated Physical AI Action-Outcome Memory on a synthetic robot/fleet edge fixture.

The experiment tests whether a context-gated temporal memory can better avoid previously risky robot actions while still choosing the right recovery action.

Hypothesis

Physical AI agents need more than semantic similarity and generic runbook priors. A past robot action is only safe to reuse when the temporal motif, topology, payload, human proximity, and change/environment context are compatible.

Context-gated Action-Outcome Memory should therefore improve:

• recovery Top-1 hit rate,
• risky-repeat avoidance,
• hazardous top-action suppression,
• auditability of misleading but retrieved prior incidents.

Inputs

• Fixture: docs/research/fixtures/physical_ai_action_outcome_episodes.json
• Harness: docs/research/tools/physical_ai_action_outcome_baseline.py
• Result: docs/research/results/physical_ai_action_outcome_013.md

Baselines

Variant	Pattern approximated	What it can miss
similar_robot_incident	Similar robot incident retrieval.	Finds superficially similar incidents without learning whether an action was safe in this context.
runbook_action_prior	Same-incident-family runbook or historical action prior.	Reuses historically successful actions even when payload, zone, human proximity, or weather differs.
reflection_only_memory	Experiential memory from reflections/postmortems.	Recalls text and outcomes, but lacks structured gates over telemetry/topology context.
context_gated_physical_ai_action_outcome	Target Physical AI Action-Outcome Memory.	Still heuristic, but explicitly gates past outcomes before action aggregation.

Procedure

1. Load five Physical AI incident families:
   • warehouse AGV dock slip,
   • mobile robot LiDAR occlusion,
   • industrial arm torque spike,
   • cold-chain temperature excursion,
   • drone GPS drift near geofence boundary.
2. For each family, evaluate one failed/unsafe query episode.
3. Include hard distractors where the unsafe action previously succeeded only under a different safety context, such as training lanes, supervised airspace, simulated payloads, or no-human cells.
4. Rank historical incidents and recommended actions for each baseline.
5. Measure whether the method avoids repeating the unsafe query action.
6. Measure whether the top action is in the expected safe recovery-action set.
7. Record top-3 retrieval quality and context-gate suppressions.

Command

python3 docs/research/tools/physical_ai_action_outcome_baseline.py \
  --fixture docs/research/fixtures/physical_ai_action_outcome_episodes.json \
  --output docs/research/results/physical_ai_action_outcome_013.md

Acceptance Criteria

Criterion	Required
Fixture parses as JSON	pass
Harness compiles with py_compile	pass
All four variants run	pass
Query families evaluated	5
Context-gated risky-repeat avoidance	1.00
Context-gated recovery Top-1 hit rate	1.00
Context-gated hazardous top-action rate	0.00
Report records suppressions	yes

Result

See docs/research/results/physical_ai_action_outcome_013.md.

Summary:

• context_gated_physical_ai_action_outcome reached 1.00 recovery Top-1 hit rate.
• It reached 1.00 risky-repeat avoidance.
• It reduced hazardous top-action rate to 0.00.
• The three non-gated baselines each recorded 0.00 recovery Top-1 hit rate, 0.00 risky-repeat avoidance, and 1.00 hazardous top-action rate on the hard distractor fixture.
• The gated variant retained misleading incidents in the evidence view for audit while suppressing unsafe outcome reuse during action aggregation.

Interpretation

The commercial wedge is not “robot incident search.” It is deciding whether a past robot outcome is safe to reuse in the current temporal context.

This maps cleanly to ZeptoDB’s Physical AI positioning: ROS/robot telemetry provides the event-time substrate, while Action-Outcome Memory records actions, outcomes, suppressions, and replayable evidence.

Next Best Step

Replay the fixture through native ZeptoDB SQL tables:

• telemetry windows,
• action/outcome tables,
• ASOF JOIN between robot state and incident/action rows,
• ROW_NUMBER ranking for candidate recommendations,
• suppression JOIN audit.

Completed by Experiment 014: docs/research/physical_ai_action_outcome_sql_replay_experiment_014.md.