AutoDrift

AutoDrift is a simulation-first engineering project for friction-adaptive autonomous drifting, emergency obstacle avoidance, and reproducible policy evaluation.

The first implementation target is deliberately narrow and practical:

• provide a reliable vehicle simulation environment with randomized road friction, vehicle mass, center-of-gravity, tire stiffness, and actuator delay;
• train and evaluate RL policies on circular drift, friction-change, and later AEB-infeasible obstacle-avoidance tasks;
• compare those policies against model-based and rule-based baselines;
• keep real hardware out of the first milestone, while keeping interfaces clean enough to add high-fidelity simulators or hardware later.

Current MVP

The repository currently contains:

• autodrift.dynamics: single-track RWD vehicle dynamics with nonlinear tire saturation and randomized physical parameters.
• autodrift.tasks: simple path tasks, starting with a circular drift track.
• autodrift.env: a Gymnasium environment for drift tracking.
• autodrift.vector_env: synchronous multi-environment rollout support.
• autodrift.config: JSON config builders for env randomization and curricula.
• autodrift.policies: baseline policies for sanity checks.
• autodrift.evaluate: command-line evaluator.
• autodrift.benchmark: shared-seed benchmark runner with CSV/JSON artifacts.
• autodrift.rollout: selected-episode traces and plots.
• autodrift.artifacts: run directory, JSON, and CSV helpers.
• docs/: literature notes and PDFs.

The project definition is in docs/implementation-plan.md. The target outcome is a complete runnable stack, not a paper prototype: simulator, training, evaluation, saved policies, benchmark reports, and documentation.

Setup details are in docs/setup.md. The default environment targets the local NVIDIA GPU stack; a CPU-only fallback is also provided.

Quick Check

pytest
PYTHONPATH=src python -m autodrift.evaluate --episodes 5 --policy heuristic
PYTHONPATH=src python -m autodrift.train_ppo --config configs/ppo_smoke.json
PYTHONPATH=src python -m autodrift.benchmark --episodes 2 --policies heuristic random

The environment intentionally hides friction by default. A learning policy must infer it from state/action history rather than reading a privileged mu value.

Training, evaluation, and benchmark commands write reproducible artifacts under runs/ by default. That directory is ignored by git; tracked configuration templates live under configs/.