--- license: apache-2.0 library_name: pytorch tags: - autonomous-driving - imitation-learning - simulation - dagger - next-lab --- ![drive](https://cdn-uploads.huggingface.co/production/uploads/67d46bc5fe6ad6f6511d6f44/hivgAahjxw3ROldCHOU-r.png) # Next Drive Next Drive is a small learned driver for a **simulated** city intersection. Every 0.1 s it chooses one of six longitudinal actions (brake hard, brake, ease off, hold speed, speed up, accelerate) while following a straight, left or right route through a traffic light, a crosswalk, a lead car that brakes suddenly, oncoming and cross traffic, and parked cars that hide a pedestrian. **Simulation and education only. It is not connected to, and has never been tested on, a real vehicle, and nothing here supports a claim about real-road safety.** **Space:** https://huggingface.co/spaces/thelamapi/Next-Drive — pick a scene, density and driver; the drive is generated in your browser and can be replayed with its critical events. ## Results on unseen test scenes (450 per driver; seeds ≥ 10000; density 3 never seen in training) | Driver | Collisions | Red-light violations | Goal reached | Hard brakes / drive | Time to goal (s) | Decision ms (mean / p95) | |---|---:|---:|---:|---:|---:|---:| | Learned + safety check | 4.4% | 0.4% | 89.1% | 1.67 | 21.0 | 0.34 / 0.47 | | Learned | 4.4% | 10.0% | 89.8% | 0.95 | 19.3 | 0.28 / 0.40 | | Rule-based | 27.1% | 0.0% | 71.1% | 0.95 | 23.3 | 0.02 / 0.07 | | Expert (teacher) | 2.7% | 1.3% | 93.6% | 1.01 | 19.2 | 1.61 / 2.38 | Collision / goal-reached rate by scene type: | Scene | Learned + safety | Learned | Rule-based | Expert | |---|---:|---:|---:|---:| | light | 4% / 96% | 2% / 98% | 1% / 99% | 1% / 99% | | crosswalk | 7% / 93% | 1% / 99% | 9% / 86% | 0% / 100% | | lead_brake | 8% / 92% | 1% / 99% | 3% / 97% | 2% / 98% | | occlusion | 2% / 79% | 10% / 73% | 68% / 32% | 7% / 82% | | mixed | 1% / 86% | 8% / 80% | 54% / 42% | 3% / 89% | Collision rate by traffic density: | Density | Learned + safety | Learned | Rule-based | Expert | |---|---:|---:|---:|---:| | 1 | 1% | 2% | 23% | 0% | | 2 | 5% | 3% | 23% | 3% | | 3 (unseen) | 7% | 8% | 35% | 5% | Rates count drives with at least one event. Decision latency was measured in Python on NVIDIA L4, single process; the Space reports its own browser timing. All 138 failures of the learned drivers are listed in `evaluation.json` and can be replayed in the Space. ![Benchmark](preview.png) [Replay video](replay.mp4): one unselected test scene, then failure replays. ## How it works * **Simulator:** 1-lane roads, NS/EW signals with yellow and all-red, IDM-like traffic, pedestrians, and line-of-sight occlusion by parked cars. The ego car perceives cars always and pedestrians only when visible within 60 m. * **Expert:** evaluates 6×6 two-phase action sequences (1 s + 2 s) at 0.25 s resolution with constant-velocity predictions, the known light schedule, a 2 m + 1.2 s following gap, and a speed cap that allows stopping before occluded gaps. * **Learned driver:** MLP 21→64→64→6 on hand-made features (speed, light state, stop-line distance, lead gap, time-to-collision under three accelerations, visible-pedestrian and occlusion distances, route). Behaviour cloning on 300 expert drives (59750 states), then 3 DAgger round(s); iteration 2 was selected on validation seeds. * **Safety check:** keeps an action only if braking at −6 m/s² after 0.5 s still avoids every predicted collision and red-light entry; otherwise picks the strongest safe deceleration. * **Rule-based baseline:** IDM car following plus stop-line, visible-pedestrian and left-turn-yield rules. * The browser runs a line-by-line JavaScript port; Node parity: 8 full drives, 0 mismatches. ## Usage ```bash python inference.py --seed 10007 --kind mixed --density 2 --driver shield python inference.py --selftest ``` `policy.safetensors` holds the PyTorch weights; `model.json` the same weights for NumPy and the browser. ## Limitations * Speed-only control on fixed routes; no lane changes, no steering errors, no sensor noise beyond occlusion. * Other road users follow simple scripted behaviour and never react to the ego car. * The expert is imperfect too (see the table); the learned driver copies its habits, including frequent hard braking at yellow lights. * All numbers come from this simulator and say nothing about real driving.