emotiv-ecot

Embodied Chain-of-Thought episodes where the body is a human cortex: one person in an EMOTIV EPOC X talking to an agent that reads a one-line brain summary before every reply. Each turn becomes a LeRobot v3.0 episode (Zawalski et al. 2024 with the robot body swapped for a head): the brain is observation and reward (Δstress, Δengagement across the reply), the agent's speech is the action, the reasoning is the per-frame TASK | AMBIENT | PLAN | TOOL | ACT | REWARD string.

An episode covers 3 s before the human speaks, the streamed answer, and 12 s of after-effect (metrics tick every ~10 s, so the tail is long enough to catch the brain's response); ambient baselines carry TASK: ambient.

Load it

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("cagataydev/emotiv-ecot")
frame = ds[80]
frame["observation.state"].shape   # torch.Size([70]), 14 ch × 5 bands
frame["task"]                      # this tick's ECoT string

No videos (use_videos=False). The v3.0 tag tracks the head commit; lerobot requires it.

Schema (fps = 8, 125 ms ticks)

key	shape	semantics
observation.state	[70]	band power, channel-major, Cortex pow order
observation.contact_quality	[14]	per-channel CQ 0 to 4
observation.motion	[10]	Q0 Q1 Q2 Q3 ACCX ACCY ACCZ MAGX MAGY MAGZ (+ motion_valid [1])
observation.metrics	[7]	attention engagement excitement longExcitement stress relaxation interest; absent → -1 (+ metrics_valid [7])
observation.facial	[6]	eye one-hot + upper/lower face power (+ facial_label [2] vocab indices)
observation.events	[12]	multi-hot: blink wink_left wink_right head_turn_left head_turn_right nod clench smile focus_high focus_low stress_high command
action	[4]	[spoke, tool_called, marker_injected, turn_length_tokens_norm]
task	string	the ECoT text (deduped in meta/tasks.parquet)

Channel order: AF3 F7 F3 FC5 T7 P7 O1 O2 P8 T8 FC6 F4 F8 AF4. Bands: theta alpha betaL betaH gamma. Slower streams resample last-known-value, faster ones latest-sample, events OR-accumulate.

REWARD is the first metric sample after the reply minus the last one before it. It reads Δstress=nan when the next slow metric tick (one every ~10 s) never arrived in the tail, and in every episode recorded before strands-emotiv 0.1.1 (their reward windows predate this fix). Absent data stays absent, never zeroed.

Collection

▶ Watch an episode recorded live (1 min)

strands-emotiv: a live dashboard mirrors the headset while a Strands agent chats with the wearer. Recording is a visible REC panel plus agent tools (record_start, record_stop, record_publish); a jaw clench vetoes consent, poor contact quality refuses recording.

Limitations

• n = 1: one brain, one headset, self-recorded. Personal research data, not a population study.
• Consumer EEG: band power from Cortex (raw EEG is license-gated), 14 saline electrodes, motion artifacts.
• The metrics (stress, engagement, …) are EMOTIV's proprietary estimates, taken as-is.
• Not medical.

Citation

@misc{emotiv-ecot,
  author = {Cagatay Cali},
  title  = {emotiv-ecot: Embodied Chain-of-Thought episodes from a human cortex},
  year   = {2026},
  url    = {https://huggingface.co/datasets/cagataydev/emotiv-ecot}
}