louisye/social-robotics-reasonable-emotion

Social Robotics: Reasonable Emotion (03b)

Was the bystander's facial-emotion trajectory contextually appropriate to the task?

One layer of the Social-Affective Filter (SAF) — dehydrated social-signal metadata extracted from egocentric (first-person) video so robots can learn to read human reactions. No raw pixels and no audio. Each row is one source video, keyed by video_id; rehydrate against your own legally-obtained Ego4D copies (below).

• Rows: 306 — videos in the evaluation slice for which this layer produced a measured signal (videos it could not measure are excluded from this per-layer dataset; the layers still join 1:1 on video_id).
• Signal: A signed −1…+1 'reasonableness' score for how the bystander's emotional response evolved during a task's reaction window: positive = emotion moved in a contextually-appropriate direction, negative = contextually inappropriate, ~0 = neutral/no change (late-stage-weighted across emotion-transition slices).
• Method: HSEmotion face emotion + an LLM (Gemma) reasoning pass over emotion transitions.

⚠️ Read this first — interpretation caveats

• Very sparse — bystander faces are rarely both visible and emotion-legible in egocentric video. Treat as illustrative, not statistically representative.
• The score is signed (−1…+1): a negative value means the emotional trajectory ran against contextual expectation, not merely a low score.
• Egocentric footage is legitimately low-yield (small/sparse bystander faces, heavy camera motion); we publish honest measurements only, never fabricated zeros.

Columns

Identity & manifest (shared across all SAF datasets)

column	type	meaning
video_id	string	Ego4D source-clip UUID. The rehydration key — map back to your own legally-obtained Ego4D copy (<video_id>.mp4).
source_dataset	string	Origin corpus (ego4d).
task_labels	string	Comma-joined VLM task label(s) — the activity the camera-wearer performed.
duration_sec	float	Source clip duration (seconds).
fps	float	Source clip frame rate.

Reasonable Emotion signal

column	type	meaning
reasonable_emotion_tasks_analyzed_raw	JSON string	Per-task detail: task_label, task_reaction_window_sec, segment_index, per-person temporal_slices ({window_sec, transition_pair, terminal_magnitude, classified_direction, slice_success_scalar} — slice_success_scalar is signed: emotion magnitude negated when classified_direction is contextually inappropriate), late_stage_weighted_success_score, task_aggregate_score (both −1…+1, signed).
reasonable_emotion_avg_task_score	float (−1…+1, signed)	Mean across tasks of task_aggregate_score — positive = contextually reasonable emotional trajectory, negative = contextually inappropriate, ~0 = neutral.
reasonable_emotion_emotion_source	string	Which emotion backend produced the underlying labels (face-emotion model vs VLM fallback).

The *_raw JSON column

The *_raw column holds the full nested per-task / per-person detail as a JSON string. Parse it with:

import json, pandas as pd
df = pd.read_parquet("hf://datasets/louisye/social-robotics-reasonable-emotion/social_metadata.parquet")
raw_col = next(c for c in df.columns if c.endswith("_raw"))
detail = json.loads(df.iloc[0][raw_col])

How to load

import pandas as pd
df = pd.read_parquet("hf://datasets/louisye/social-robotics-reasonable-emotion/social_metadata.parquet")
# or:  from datasets import load_dataset;  ds = load_dataset("louisye/social-robotics-reasonable-emotion")

Rehydration — mapping back to video

video_id is the Ego4D clip UUID. With your own licensed Ego4D copy, the file is <video_id>.mp4; timestamps in the *_raw columns index into that clip. A helper (rehydrate_dataset.py) is included. We never redistribute source media — obtain Ego4D under its own license.

Provenance

Generated by the SAF pipeline (export_metadata.json records schema_version + pipeline_git_sha). Headers are the descriptive layer name + metric; the pipeline's internal 03b_ layer-id prefix is stripped at publish time. License MIT (this metadata only; Ego4D videos remain under the Ego4D license).