BAAI/Discoverse-L

This dataset is the result of joint research conducted by Hao Tang's team at the School of Computer Science, Peking University, and the Beijing Academy of Artificial Intelligence (BAAI).

Paper: EvoVLA: Self-Evolving Vision-Language-Action Model Authors: Zeting Liu*, Zida Yang*, Zeyu Zhang*†, Hao Tang‡
Institution: Peking University

Overview

Discoverse-L is a long-horizon manipulation benchmark built on the DISCOVERSE simulator with AIRBOT-Play robot platform. It provides:

• 3 multi-stage manipulation tasks with varying difficulty:

  • Block Bridge (74 stages): Place two bars to form a bridge structure, then fill with multiple blocks
  • Stack (18 stages): Stack three colored blocks in sequence
  • Jujube-Cup (19 stages): Place a jujube fruit into a cup and move the cup onto a plate

• 50 scripted demonstration trajectories per task (150 total)

• Task-aligned normalization statistics for cross-task training

• Stage dictionaries with Gemini-generated triplets (positive, negative, hard-negative text descriptions)

Dataset Structure

Discoverse-L/
├── demonstrations/
│   ├── block_bridge_place/
│   │   ├── 000/
│   │   │   ├── obs_action.json    # Joint states & actions
│   │   │   ├── cam_0.mp4           # Main camera view
│   │   │   └── cam_1.mp4           # Wrist camera view
│   │   ├── 001/
│   │   └── ... (50 episodes)
│   ├── place_jujube_coffeecup/
│   │   └── ... (50 episodes)
│   └── stack_block/
│       └── ... (50 episodes)
├── metadata/
│   └── task_aligned_normalization.json  # q01/q99/mean/std for each task
└── stage_dictionaries/
    ├── block_bridge_place_stages.json
    ├── place_jujube_coffeecup_stages.json
    └── stack_block_stages.json

Data Format

Demonstration Trajectories

Each episode directory contains:

• obs_action.json: Time-aligned observations and actions

  {
    "time": [t1, t2, ...],
    "obs": {
      "jq": [[q0, q1, q2, q3, q4, q5, q6], ...]  // Joint positions
    },
    "act": [[a0, a1, a2, a3, a4, a5, a6], ...]   // Joint actions
  }
  

• cam_0.mp4: 448×448 main camera view (20 FPS)
• cam_1.mp4: 448×448 wrist camera view (20 FPS)

Task-Aligned Normalization

Computed from the 50 demonstrations per task:

{
  "task_name": {
    "action": {
      "mean": [7-dim],
      "std": [7-dim],
      "min": [7-dim],
      "max": [7-dim],
      "q01": [7-dim],  // 1st percentile
      "q99": [7-dim],  // 99th percentile
      "mask": [bool×7] // True for continuous joints, False for gripper
    }
  }
}

Stage Dictionaries

Gemini-2.5-Pro generated text triplets for each stage:

[
  {
    "id": 0,
    "positive": "The robotic gripper is approaching the target object",
    "negative": "The gripper is moving away from all objects",
    "hard_negative": "The gripper is grasping a distractor object"
  },
  ...
]