oscarqjh/BEHAVIOR-1K_easi

BEHAVIOR-1K EASI Dataset

Episode definitions and all required assets for the BEHAVIOR-1K 2025 Challenge, packaged for use with the EASI evaluation framework.

Splits

Split	Episodes	Description
b50_test_public	1,000	50 tasks × 20 public test instances
b50_train	~10,000	50 tasks × ~200 training instances (no test overlap)

The B10 sub-split (first 10 tasks × 20 instances = 200 episodes) is derived from b50_test_public via a max_task_idx: 9 filter in the EASI task YAML.

Usage with EASI

easi task download behavior1k_b50_test_public   # downloads + extracts all zips
easi start behavior1k_b50_test_public --agent react --backend openai --model gpt-4o

Files

File	Size	Description
episodes_b50_test_public.jsonl	~200 KB	1,000 episode definitions
episodes_b50_train.jsonl	~2 MB	~10,000 episode definitions
omnigibson.key	44 B	Decryption key for encrypted USD assets
behavior-1k-assets.zip	~33 GB	Scene + object USD models (git LFS)
omnigibson-robot-assets.zip	~2.4 GB	Robot USD models (git LFS)
2025-challenge-task-instances.zip	~400 MB	TRO state files + metadata (git LFS)

Episode Schema

{
  "episode_id":             "turning_on_radio__242",
  "task_name":              "turning_on_radio",
  "task_idx":               0,
  "scene_model":            "house_double_floor_lower",
  "activity_definition_id": 0,
  "instance_id":            242,
  "max_steps":              3912,
  "human_length":           1956.0,
  "human_distance":         1.2958,
  "human_left_eef":         3.5632,
  "human_right_eef":        5.328
}

Asset Extraction

EASI automatically extracts the three zip files on first download. After extraction, OMNIGIBSON_DATA_PATH is set to the local dataset directory so OmniGibson finds all assets (scenes, objects, robots) at the standard paths.

Requirements

• OmniGibson v3.7.2 + Isaac Sim 4.5.0 (easi env install omnigibson:v3_7_2)
• CUDA 12.4