donkeycar-rl-autoresearch/agent
Paul Huliganga 2d52bb4ffc fix(core): replace exploit bandaids with solid physics barriers + clean reward
Root cause: barriers were zero-thickness MeshCollider planes with no CCD on the
car. The car tunnelled through between frames. Every Python patch was trying to
catch in code what physics should enforce.

Unity (source only — build in progress):
- RoadBuilder.cs: CreateBarrier() now makes BoxCollider-per-segment with real 3D
  volume (barrierThickness=1.0m default) + half-thickness overlap at corners to
  seal gaps. CreateEndCap() seals open ends of non-looping tracks (generated_road).
- Car.cs: rb.collisionDetectionMode = Continuous in Awake() — prevents tunneling.

Python:
- reward_wrapper.py v7: removed CTE-patience termination, high-CTE negative
  reward, solid_hit monitoring, low-speed/wedge detection. Kept: efficiency gate,
  no-progress (active_node) termination, lap exploit guard. Reward = speed×CTE_quality.
- exp23_generated_road_clean.py: single track, no warm-start, 200k steps, clean
  reward, MAX_EPISODE_SECONDS=120 as safety net only.
- tests: 17 tests covering clean reward properties.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-05-05 15:56:00 -04:00
..
__pycache__ Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
bin Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
experiments fix(core): replace exploit bandaids with solid physics barriers + clean reward 2026-05-05 15:56:00 -04:00
models chore(exp22): update wedgefix run log — training stopped for strategy rethink 2026-05-05 15:36:18 -04:00
outerloop-results feat: add exp17 parallel DummyVecEnv 450k training + strategy docs 2026-04-28 02:42:20 -04:00
sb3-models Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
sessions Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
test-results docs: Exp10 eval results — total failure, crashes on all tracks (massive regression from Exp9/W4T9) 2026-04-19 10:19:16 -04:00
AUTORESEARCH_README.txt AUTORESEARCH: Full Karpathy-style GP+UCB meta-controller, clean base data, fixed all paths, ready to run 2026-04-13 00:52:00 -04:00
README.md Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
SESSION_HANDOFF.md fix(core): replace exploit bandaids with solid physics barriers + clean reward 2026-05-05 15:56:00 -04:00
SETUP_QUICKSTART.md Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
TROUBLESHOOT.md Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
analysis_circular_driving.py fix: path-efficiency reward (v3) defeats circular driving exploit 2026-04-13 13:36:17 -04:00
auth.json Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
autoresearch Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
autoresearch_controller.py milestone: Phase 1 complete — genuine driving confirmed; launch Phase 2 corner learning 2026-04-13 19:33:06 -04:00
behavioral_wrappers.py feat: Phase 3 — behavioral control, enhanced evaluator, 53 tests 2026-04-14 09:28:43 -04:00
champion_tracker.py feat: Wave 1 complete — real PPO training, model save, GP+UCB autoresearch, 37 tests passing 2026-04-13 10:03:15 -04:00
check_envs.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
choose_and_run_track.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
discretize_action.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
donkeycar_autoresearch.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
donkeycar_outer_loop.py AUTORESEARCH: Full Karpathy-style GP+UCB meta-controller, clean base data, fixed all paths, ready to run 2026-04-13 00:52:00 -04:00
donkeycar_sb3_runner.py fix: reduce timesteps to 1k-5k for Phase 1 CPU training; add sim health/stuck detection; fix PPO throttle clamp 2026-04-13 11:17:08 -04:00
donkeycar_sb3_runner.py.README.txt Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
eval_on_track.py feat: v5 reward — speed × CTE-quality, drop efficiency term 2026-04-17 13:25:38 -04:00
evaluate_champion.py fix: track switching via unwrapped viewer.exit_scene() — automatic scene changes work 2026-04-14 10:04:15 -04:00
list_tracks.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
manual-multiepisode-2.log Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
manual-multiepisode-3.log Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
manual-multiepisode-batch.log Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
manual-multiepisode.log Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
manual_multiepisode-batch-run.log Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
manual_multiepisode_batch.sh Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
multitrack_eval.py feat: comprehensive multi-track evaluation script + research log updates 2026-04-14 10:11:47 -04:00
multitrack_runner.py fix: exp19 — hard episode time limit to stop minutes-long stuck cars 2026-04-28 09:18:04 -04:00
park_at_start.py wave3: add multi-track autoresearch system (83 tests passing) 2026-04-14 12:47:12 -04:00
reward_wrapper.py fix(core): replace exploit bandaids with solid physics barriers + clean reward 2026-05-05 15:56:00 -04:00
run_all_known_tracks.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
run_circuit_launch.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
run_donkeycar_test.sh Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
run_eval.py fix: reward v6 — efficiency gate prevents circular driving, stuck_steps 80→40 2026-04-19 12:02:55 -04:00
settings.json Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
test_donkeycar.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
test_donkeycar_gymnasium.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
test_sdsandbox.py Initial commit: stable RL sweep runner, legacy and new scripts, full docs included 2026-04-12 22:57:50 -04:00
tmp_eval_mountain_candidates.py docs: capture robust mountain finetune winner at 36k and preserve eval comparison 2026-04-20 00:43:27 -04:00
track_switcher.py fix: track switching via unwrapped viewer.exit_scene() — automatic scene changes work 2026-04-14 10:04:15 -04:00
wave3_controller.py fix: complete LR override — must patch lr_schedule, not just param_groups 2026-04-14 21:27:43 -04:00
wave4_controller.py fix: prevent trial timeouts losing all data 2026-04-15 21:54:50 -04:00

README.md

DonkeyCar RL Batch Automation and Setup Guide

System Requirements

  • Windows 10/11 machine (with DonkeyCar Unity Simulator installed)
  • WSL2 with Ubuntu (Python 3.x installed)
  • DonkeyCar Unity Simulator running in Windows, Remote Control enabled

1. WSL (Linux) Setup

A. Install Python, pip, and core packages

sudo apt update && sudo apt install -y python3-pip git
pip3 install --upgrade pip
pip3 install stable-baselines3 gymnasium gym-donkeycar numpy matplotlib
# If pip install gym-donkeycar fails, use:
pip3 install git+https://github.com/tawnkramer/gym-donkeycar.git

B. Place scripts

Copy these into /home/paulh/.pi/agent/:

  • donkeycar_sb3_runner.py
  • manual_multiepisode_batch.sh (and make executable)
  • Any grid/outer loop script (donkeycar_outer_loop.py, as needed)

2. Unity DonkeyCar Simulator Setup (Windows)

  • Download/install Unity DonkeyCar sim (from DonkeyCar/tawnkramer Github or releases page)
  • Open simulator, select "Donkey Generated Track" and enable Remote (SocketAPI) mode
  • Ensure port 9091 is listening (default); leave sim running and visible

3. Running Robust Batches (Best Practice)

Clean multi-episode approach:

Do NOT rapidly kill/restart agents. Use scripts that:

  • Open one connection, run multiple episodes using env.reset() in a loop
  • Cleanly call env.close()
  • Batch script launches process, waits a couple seconds, repeats

Sample batch script (manual_multiepisode_batch.sh):

#!/bin/bash
for i in {1..20}; do
  echo "===== RUN $i ===== $(date)" | tee -a manual-multiepisode-batch.log
  python3 donkeycar_sb3_runner.py >> manual-multiepisode-batch.log 2>&1
  echo "===== END RUN $i ===== $(date)" | tee -a manual-multiepisode-batch.log
  sleep 2
done

Make executable:

chmod +x manual_multiepisode_batch.sh

Run in background:

nohup ./manual_multiepisode_batch.sh &

4. Debugging/Recovery

  • If sim blanks/hangs, first check for stuck agent processes: ps aux | grep donkeycar_sb3_runner
  • Always allow scripts to call env.close()
  • If hang persists: restart Unity DonkeyCar Sim in Windows

5. Automation Flow Summary

Step Where Details
Install WSL/Linux pip install ...
Prepare WSL/Linux Place scripts, ensure Python dependencies
Sim start Windows Start DonkeyCar Unity Sim
Run batch WSL/Linux nohup ./manual_multiepisode_batch.sh ...
Monitor Both Car resets, batch log grows, check hangs

See included Python scripts for reusable RL runner and grid search outer loop logic.