How anonymous-report-421/gpt-as-policy works
This repository evaluates GPT 6 Astra as an embodied policy, comparing direct GPT actions with a hybrid policy that reviews π₀.₅ proposals, and publishes a bilingual technical report. The principal runtime path is case setup, simulator observation, policy proposal and review, action execution, then recorded results. Separate RoboLab integrations and local annotation/review tools support related evaluation work. The report frontend and preview serve a distinct publication workflow. The README documents the overall capabilities; several subsystem links below are dashed where the sampled source does not establish exact wiring.
Main entry points: client.py Rollout Client · robodojo_server/server.py Simulator Server · robolab_server/server.py RoboLab Server · main.jsx Report Frontend
Components 24
RoboDojo Evaluation
- Rollout Tools
hybrid_rollout/robodojo/skill/run.py - Rollout Client
hybrid_rollout/robodojo/robodojo_server/client.py - Simulator Server
hybrid_rollout/robodojo/robodojo_server/server.py - Episode Session
hybrid_rollout/robodojo/robodojo_server/session.py - Case Evaluation
hybrid_rollout/robodojo/evaluation.py - Run Artifacts
hybrid_rollout/robodojo/io.py
Policy Services
- π₀.₅ Service
hybrid_rollout/robodojo/pi05_server/server.py - π₀.₅ Client
hybrid_rollout/robodojo/pi05_server/client.py - GPT Baseline
hybrid_rollout/robodojo/robodojo_server/gpt_only_client.py - Decision Records
hybrid_rollout/robodojo/robodojo_server/decision_log.py
RoboLab Integration
- RoboLab Server
hybrid_rollout/robolab/robolab_server/server.py - RoboLab Session
hybrid_rollout/robolab/robolab_server/session.py - EEF Controller
hybrid_rollout/robolab/robolab_server/eef_controller.py - RoboLab π₀.₅
hybrid_rollout/robolab/pi05_server/server.py
Report Experience
- Report Frontend
hybrid_rollout/report_site/app/src/main.jsx - Report Content
hybrid_rollout/report_site/app/src/content/report/ReportContent.jsx - Chart State
hybrid_rollout/report_site/app/src/charting/ChartState.jsx - Query Boundary
hybrid_rollout/report_site/app/src/components/QueryDataBoundary.jsx - Local Preview
hybrid_rollout/report_site/preview.py - Report Server
hybrid_rollout/report_server.py
Annotation Review
- Annotation Dashboard
hybrid_rollout/annotation_dashboard.py - Annotation UI
hybrid_rollout/annotation_ui/app.js - History Dashboard
hybrid_rollout/history_dashboard.py
Other components
- Evaluator
Connections 22
- Evaluator → Rollout Tools: starts rollout
- Rollout Tools → Rollout Client: drives
- Rollout Client → Simulator Server: RPC calls
hybrid_rollout/robodojo/robodojo_server/client.py - Simulator Server → Episode Session: serves episodes
hybrid_rollout/robodojo/robodojo_server/server.py - Simulator Server → Case Evaluation: validates cases
hybrid_rollout/robodojo/robodojo_server/server.py - Rollout Client → π₀.₅ Client: requests inference
- π₀.₅ Client → π₀.₅ Service: WebSocket inference
- Rollout Client → GPT Baseline: alternative baseline
hybrid_rollout/robodojo/robodojo_server/gpt_only_client.py - Rollout Client → Run Artifacts: writes run data
hybrid_rollout/robodojo/robodojo_server/client.py - Rollout Client → Decision Records: records decisions
hybrid_rollout/robodojo/robodojo_server/client.py - Evaluator → RoboLab Server: runs RoboLab
- RoboLab Server → RoboLab Session: creates session
hybrid_rollout/robolab/robolab_server/server.py - RoboLab Server → EEF Controller: uses controller
- Evaluator → RoboLab π₀.₅: runs policy service
- Report Frontend → Report Content: renders
- Query Boundary → Chart State: shows loading state
hybrid_rollout/report_site/app/src/components/QueryDataBoundary.jsx - Report Frontend → Query Boundary: hosts query views
- Evaluator → Local Preview: previews report
README.md - Local Preview → Report Server: serves report
- Evaluator → Annotation Dashboard: reviews annotations
- Annotation Dashboard → Annotation UI: serves interface
- Evaluator → History Dashboard: inspects history