Skip to content

INDEX.md

Latest commit

 

History

History
75 lines (65 loc) · 5.54 KB

INDEX.md

File metadata and controls

75 lines (65 loc) · 5.54 KB
Raw

AutoDrift — navigation index

The repo is large because the autonomous research process emits a load-bearing artifact tree (docs/m####-*.md milestone records + generated src/autodrift/*.py + experiments/manifests/*.json), enforced on every commit by python -m autodrift.research_validate. This index points to the human-meaningful deliverables and findings amid that scaffolding.

Claim boundary — what is canonical vs superseded (read first)

The repo spans two eras and the claim boundary moved; to avoid confusion:

  • CANONICAL (current, defensible):
    • Bounded two-regime conclusion (two-regime-thesis-drift-2026-06.md): deliberate drift is unnecessary before slip under the force-envelope theorem and M3270 finite Chrono certificate. After slip, expanded controls imply only weak recovery-set inclusion; M3271-M3273 found no strict steering witness.
    • Foundation audit (foundation-audit-drift-required-label-2026-06.md): the drift_required label was built on a ~2x-wrong grip assumption; true must-drift fraction ~2%.
    • F2 conditional RL result (paper/c5prime/main.tex, m3264-*.md): one policy beats the reflex floor + scripted oracle on the saturated drift cell (0.856) but regresses on benign avoidance (0.700). Engineering-only, not a promotion. This is the headline scientific claim.
  • SUPERSEDED (record, not a live claim):
    • "Drift-to-avoid beats rule-based" (drift-to-avoid-driver-2026-06.md, the 0.93-vs-AEB number) — premise falsified by the foundation audit.
    • "Strongest/most-general driver certified" / "one policy for the full passenger-car fleet" (the pre-takeover capstone + the C5 spread formulation) — the universal-driver spread claim was pricing-rejected (0/8 cells), Track C closed (current-status.md); the surviving result is the scoped F2 conditional finding above, not a fleet-wide solution.

Core simulation + RL code (src/autodrift/)

  • dynamics.py — single-track RWD vehicle, nonlinear tire saturation (friction-circle capped), DR params.
  • env.py — Gymnasium env (obs72, friction step, obstacle task + scenario labels), ObstacleTaskConfig.
  • scenarios.py — obstacle feasibility labels (aeb/aes/drift_required/unavoidable). NOTE: the default conventional_lateral_mu_fraction=0.42 understates real grip ~2x — see the foundation audit.
  • policies.py — baselines: aeb, aes_heuristic, envelope_aes (PRIVILEGED, invalid baseline), honest_aes, mu_aware_aes (mechanism probe).
  • vector_env.py, config.py, train_ppo.py, benchmark.py, evaluate.py — training/eval harness.
  • gpu_env*.py, gpu_physics_pwr*.py, gpu_sim/ — fast GPU surrogate (pwrBD = faithful longitudinal config).

Trained drivers (checkpoints under runs/feasibility_audit/phase4_f2/)

  • distill_final_robustified_capstone.pt — the certified one-hot FiLM driver (drift+avoid spectrum x 3 vehicles, 132/136 seed-clustered CI). Needs a vehicle one-hot label.
  • selfid_fullscenario_best.pt — LABEL-FREE self-ID driver (infers vehicle from history; drift 1.0 + avoid 0.956).
  • selfid_rma_round1.pt — avoid-only RMA self-ID (0.970, A-comparable).
  • runs/ppo_fullspectrum/fullspectrum_avoid_driver.pt — full-spectrum avoidance PPO driver.

Key findings (docs/, 2026-06)

  • two-regime-thesis-drift-2026-06.md — bounded synthesis: no deliberate drift before slip in the theorem/M3270 scope; after slip, extra steering value is conditional and was not strict in the corrected panels.
  • foundation-audit-drift-required-label-2026-06.md — the drift_required label is built on a 2x-wrong grip assumption; 5 measurements (planar + Chrono) + adversarial verification show drift gives no avoidance advantage.
  • honest-avoidance-driver-2026-06.md — the RL avoidance driver's real differentiator is closed-loop limit control (not drift, not mu-knowledge); fixed rules slide out even with true mu.
  • selfid-rma-experiment-2026-06.md — replacing the vehicle one-hot with history-inferred RMA self-ID; recovers avoid + generalises to unseen vehicles (leave-one-out 0.965).
  • coverage-spectrum-design-2026-06.md — the S1/S2/S3 spectrum, cross-vehicle, certification, perception-DR.
  • multi-fidelity-gpu-rewrite-design-2026-06.md, m5-emergency-avoidance.md — GPU rewrite + M5 obstacle taxonomy.

Reproducible audits (scripts/audits/)

  • measure_env_lateral_capacity.py — env's real conventional vs drift lateral force capacity.
  • drift_reachability.py / box_reachability.py / box_reachability_angled.py — CG + oriented-box avoidance reachability (drift gives no edge, even angled/extended).
  • chrono_drift_reachability.py — faithful-Chrono confirmation.
  • recovery_reachability.py / chrono_recovery.py — deprecated invalid audits retained for provenance; wrong pedal normalization and no production ESC semantics. Use M3271-M3273.

Session experiment scripts (scripts/feasibility_audit/)

185 files — the F2/capstone/self-ID pipeline (e.g. distill_both_final_integrated.py, distill_both_3vehicle_film.py, phase4_f2_train.py, selfid_models.py, _selfid_*.py, _s3_*.py). The _-prefixed ones are one-off analysis/validation scripts for specific milestones.

Process scaffolding (do not hand-edit)

  • docs/m####-*.md (3263) + experiments/manifests/*.json (3183) + experiments/research_queue.csv — the autonomous research-process state, required by the research_validate commit hook. scratch/ (gitignored) holds throwaway debug probes. runs/ (gitignored, ~28GB) holds training/eval artifacts.