Policy Replay

Select Robot

CONNECTING
Stack -
Decisions -
Distribution -
Sim Time -
Runs
Live, fixture, and imported ingest sessions grouping policy decisions
Recent Runs
Run Detail
Platform Health
API, store, workers, renderer, artifact storage, and retention status
Replay Jobs
Worker queue, attempts, and review history
Worker Queue
Durable substrate for scans, rerolls, imports, and long-running platform jobs
Policy Bundle Import
Bring local SB3 checkpoints or training bundles into the compatibility catalog
Source
Path or ref
Catalog id
Checkpoint zip metadata
Owned Policy Distribution
Training-bundle import path for mean, std, log probability, entropy, and alternatives
Shadow Policy Comparison
Compare an imported policy bundle against baseline/counterfactual actor behavior and selected deployment action shape
Policy Promotion Gate
Candidate-vs-baseline release gate over replay-suite evidence and imported-policy distribution deltas
Baseline
Candidate
Suite
Counterfactual
Value
Horizon
Mean L2 Limit
Promotion Runs
Gate Result
Policy Input Attribution
Finite-difference ranking of imported-policy observation dimensions that move actor outputs
Decision Intelligence
Auto-surface surprising decisions and launch replay-backed fragility sweeps
Surprising Decisions
Fragility Candidates
Counterfactual Fragility Search
Evidence-tiered scan across decisions with policy-only and physics-backed rows
Scope
Parameter
Decisions
Top-K Reroll
Horizon
Recent Scans
Fragility Index
Improvement Flywheel
Compile a suite from the fragility index, retrain on the fragile pool, and gate the candidate against the deployed baseline
Top-N Fragile
Timesteps
Q Re-distill
Recent Flywheel Runs
Evaluation Workbench
Counterfactual sweeps, ranked candidates, and reviewable replay evidence
Min
Max
Count
Horizon
Rank By
Recent Evaluations
Ranked Candidates
Replay Suites
Repeatable G1 release gates over behavioral metrics, evidence confidence, and artifacts
Suite
Horizon
Decision Limit
Max Move Delta
Max Yaw Delta
Suite Runs
Release Gate Result
Run Timeline
Decisions, replays, evaluations, and incidents in one review stream
Artifact Catalog
Replay manifests, videos, clips, and generated frame evidence
Data Ingestion
Import ROS2-exported JSONL captures with topic counts, missing fields, and source manifests
Evidence Search
Find decisions, jobs, evaluations, artifacts, and incidents
Incident Review
Create an evidence packet from the selected decision, replay, or evaluation
Policy Lineage
Policy identity, decision volume, robot coverage, and distribution readiness
Lineage Chain
Dataset Curation
Freeze selected decisions, replays, evaluations, and incidents into review datasets
Fleet Health
Robot status derived from replay jobs, incidents, artifacts, and validation evidence
Alerts and SLOs
Rule previews for failures, high replay deltas, incidents, and missing artifacts
Edge Agent Status
Local collector capabilities, queued work, artifact counts, and offline limits
Safety-Gated Teleop
Dry-run request review only; no robot command leaves the platform
No teleop command execution is implemented in this phase.
Simulation Backend Registry
Available and future physics backends for replay, rendering, and synthetic evidence

Decisions

Decision Detail

-
Policy
-
Observation
-
Replay State
-
Counterfactual
-
Chosen Action
Actor Distribution
-
Cognition — Estimated Returns
-
Baseline Video
not captured
Counterfactual Video
needs tier 3
Tier 3 Capture Gate
blocked
Captured Telemetry
telemetry
Counterfactual Preview
policy-only
Baseline - Delta -
Policy command -