Spaces:
Sleeping
Sleeping
Commit History
-
feat: incident-response reward system + live incident management dashboard
Aksh Parekh +1: Claude Sonnet 4.6Replace diffuse per-step shaping with incident-response grading: - Every step classifies scene into: CRASH_IMMINENT, NEAR_MISS_AHEAD, NEAR_MISS_SIDE, BLOCKED_AHEAD, APPROACHING_GOAL, CLEAR_ROAD - Each incident type has a grade table: correct response = +3..+8, wrong response = -2..-15 (vs old +0.4 always-alive baseline) - Gradient is now large and directional — tied to real decisions - Policy must learn 'brake/lane_change when blocked' to get high reward Dashboard redesign: - Live incident feed: every step shows incident type, decision, reward, correct? - Response accuracy % metric (correct decisions / total) per episode - Second chart: accuracy history per episode (0–100%) - Episode log now shows accuracy column - Goal marker visible on road canvas overflow_gym_env.py: - Passes ego_x, ego_lane, decision, cars, goal_x to compute_reward() - info dict now includes incident_type, decision, grade_desc - Reads actual goal_position from env._cars[0].goal_position reward.py: - Incident…68d6c60 -
feat: continuous PPO training + live web dashboard
Aksh Parekh +1: Claude Sonnet 4.6- Replace server/app.py with FastAPI continuous training loop - PPO training runs in background thread (no manual stepping) - Live dashboard: 2D road canvas on left, reward chart + episode log on right - SSE + polling for real-time updates (0.5s heartbeat) - Capped reward mode: min(reward, 2.0) per step - Uncapped reward mode: base + token_count * 0.001 (frontier LLM scaling) - Toggle capped/uncapped via UI button or POST /api/mode - Curriculum stage display with pip indicators - PPO losses (pg/vf/entropy) shown live - Dockerfile cleaned: removed openenv-core dep to avoid websockets conflict Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>71efc66