ecot-spec

cagataycali/embodied-chain-of-thought (ECoT) — MIT, ⭐3, created 2026-06-19, spec v0.1 draft. Dataset SPECIFICATION + reference tooling for robot learning. Binds free-form agent reasoning traces ("why") to a dense robot-control corpus ("what") — one canonical store, two views: (1) VLA view = raw LeRobot v3 (front video + state + action @ fps) for OpenVLA/π0/GR00T/SmolVLA; (2) ECoT view = ChatML with native <tool_call> tokens + frame references (NOT duplicated base64), aligned step-by-step. Training target: given frames up to t + reasoning-so-far, predict next reasoning step AND next action chunk. Frame binding: frame_index = round((wall_clock_ts − episode_start_ts) × fps). Canonical SQLite schema (reasoning_events.sql, WAL, multi-writer). JSON Schema (ecot_sample.schema.json). ChatML render with native tool-call tokens (Qwen3/Llama 3.1 templates). LeRobot v3 dense spine. spec/ = overview, event taxonomy, frame binding, multi-agent, ChatML view, conformance, optional multimodal memory (CLIP/YOLO/ASR). Multi-agent: continuous concurrent reasoning contributions. Prior art: RLDS/Open-X-Embodiment, Zawalski et al. 2024 ECoT, LeRobot v3, HF tool-enabled chat templates. Reference impl lives in scout-the-rover (tools/reasoning_log.py writer + tools/ecot_export.py exporter). Author Çağatay Çalı.