Benchmark long term hazard graph memory for embodied planning

Source: Safe Task Planning with Long-Term Graph Memory for Embodied Agents (arXiv:2609.08444, submitted 2026-09-08, CoRL 2026).

The originating team reports that a persistent semantic graph of objects and relations, combined with risk prediction and conservative replanning, improves safe success under partial observability where hazards are outside the current field of view. No RuView performance claim is made until reproduction.

RuV experiment

Coordinate with ruvnet/metaharness#301 Track B. Compare:

1. local field-of-view planning
2. current RuView and WorldGraph spatial memory
3. deterministic long-term hazard graph
4. model risk predictor on the same graph

Freeze rooms, held-out environments, robot or simulator platform, action set, object-relation schema, seeds, evaluator, model route, sensor degradation, and budget.

Report safe success, ordinary task completion, missed hazards outside view, false hazards, unnecessary replans, memory bytes, update latency, planning latency, tokens, energy where measurable, and failures under stale, partial, malformed, or missing observations.

Gate

At least 30 percent fewer unsafe actions or 10 absolute points higher safe success, ordinary completion within 3 absolute points, p95 planning overhead below 15 percent, deterministic rollback, and zero authority inferred from graph confidence.

Graph state is evidence only. RVM remains the execution authority. No production routing or model deployment in this issue.

Coordinate with ruvnet/metaharness#301 Track B and first review the current RuView and WorldGraph planning and memory behavior. Define the four comparison conditions with the listed frozen variables and collect the requested safety, completion, latency, memory, token, energy, and failure metrics. Done means the stated safety or safe-success gate is met without exceeding completion or p95 overhead limits, with deterministic rollback and no authority inferred from graph confidence.