DavidxWang/worldarena2026-robotwin-data

WorldArena-RoboTwin Dataset

Dataset: DavidxWang/worldarena2026-robotwin-data
Provider: davidwang (WorldArena Challenge 2026 Track 1 team)
Last Updated: 2026-05-04
Downstream Model: ABot-PhysWorld A2V VACE Training

---

Overview

This dataset provides action-conditioned robot manipulation videos synthesized by RoboTwin 2.0 simulation, formatted specifically for ABot-PhysWorld A2V (Action-to-Video) VACE training. It covers all 50 task categories from the CVPR 2026 WorldArena Challenge Track 1 test set.

---

Dataset Scale

Split	Episodes	Videos	Actions	Total Size
episodes/ (generated)	509	509 × video.mp4	509 × actions_16d.npy	~187 MB
official_episodes/ (official test reference)	1000	—	1000 × actions.npy	~55 MB
Total	1509	509	1509	~244 MB

---

Directory Structure

training_data/
├── episodes/                          # 509 generated training episodes
│   ├── {task}__{episode_id}/          # e.g., beat_block_hammer__episode0/
│   │   ├── video.mp4                  # 640×480, 121 frames, ~300–700 KB
│   │   ├── first_frame.jpg            # 640×480 RGB JPEG, ~26 KB
│   │   ├── actions.npy                # (121, 14) float32 — joint-space actions
│   │   ├── actions_16d.npy            # (121, 16) float32 — end-effector actions ← A2V training input
│   │   ├── actions_full.npz           # Full breakdown (see below)
│   │   └── metadata.json              # Task info + 3 official-style instructions
│   └── ...
│
├── official_episodes/                 # 1000 official WorldArena test episodes
│   ├── episode{N}/                    # e.g., episode1/
│   │   ├── first_frame.png            # 240×320 RGB PNG (official resolution)
│   │   ├── actions.npy                # (N_frames, 14) float32 — joint-space actions
│   │   └── metadata.json             # Task name + 3 official instructions
│   └── ...
│
├── camera_params/                     # Shared camera parameters (all episodes use the same camera)
│   ├── head_intrinsic_params.json     # Camera intrinsics (Large_D435)
│   └── head_extrinsic_params.json    # Camera extrinsics — 121 frames, camera-to-world (c2w)
│
├── metadata_train_a2v.jsonl           # Training index for ABot A2V (509 entries)
└── manifest.json                      # Full episode manifest

---

File Format Details

actions_16d.npy — Primary A2V Training Input

Shape: (121, 16), dtype: float32

This is the key action file for ABot-PhysWorld A2V VACE training. It encodes both arms' end-effector poses in Cartesian space, which the VACE module projects onto the image plane to generate 2D trajectory maps.

Dimension layout (per frame):
  [0:3]   left_xyz          — left end-effector position (meters, world frame)
  [3:7]   left_quat_xyzw    — left end-effector rotation (quaternion, xyzw convention)
  [7]     left_gripper      — left gripper openness × 120  (0 = closed, 120 = fully open)
  [8:11]  right_xyz         — right end-effector position (meters, world frame)
  [11:15] right_quat_xyzw   — right end-effector rotation (quaternion, xyzw convention)
  [15]    right_gripper     — right gripper openness × 120

Gripper scaling note: Values are pre-multiplied by 120 to match the colormap normalization in get_traj_maps() (normalized = gripper / 120). The VACE trajectory visualization uses a green colormap for the left arm and red for the right arm, with color intensity encoding gripper openness.

Example values (episode0, frame0):

left_xyz:   [-0.298, -0.314,  0.942]  m
left_quat:  [ 0.700,  0.000,  0.000,  0.714]  (xyzw)
left_grip:  120.0  (fully open)
right_xyz:  [ 0.306, -0.313,  0.941]  m
right_quat: [ 0.704,  0.000,  0.000,  0.711]  (xyzw)
right_grip: 120.0  (fully open)

---

actions.npy — Joint-Space Actions (reference only)

Shape: (121, 14), dtype: float32

[0:6]   left arm joint angles  (radians, 6-DOF)
[6]     left gripper            (0 = closed, 1 = open)
[7:13]  right arm joint angles  (radians, 6-DOF)
[13]    right gripper           (0 = closed, 1 = open)

Not used in A2V training. Included for completeness and potential forward-kinematics experiments.

---

actions_full.npz — Full Action Breakdown

Contains 7 arrays from the same 121-frame episode:

Key	Shape	Description
joint_action_vector	(121, 14)	Same as actions.npy
joint_action_left_arm	(121, 6)	Left arm joints only
joint_action_right_arm	(121, 6)	Right arm joints only
joint_action_left_gripper	(121,)	Left gripper 0/1
joint_action_right_gripper	(121,)	Right gripper 0/1
endpose_left_endpose	(121, 7)	Left EEF pose [x,y,z,qx,qy,qz,qw] — source of actions_16d
endpose_right_endpose	(121, 7)	Right EEF pose — source of actions_16d

---

video.mp4 — Ground Truth Video

• Resolution: 640 × 480 (W × H)
• Frames: 121 (uniformly sampled from 288-frame raw simulation)
• Format: H.264, ~24 fps
• Content: Head-camera RGB view of dual-arm manipulation, rendered with Vulkan ray tracing on NVIDIA H100

---

first_frame.jpg — Condition Image

• Resolution: 640 × 480
• Format: JPEG, quality 95
• Content: Frame 0 of the corresponding video.mp4 (RGB correct, verified)
• Role in training: Passed as input_image to the I2V (Image-to-Video) pipeline

---

metadata.json — Task & Instruction Info

{
  "task": "beat_block_hammer",
  "episode_id": 0,
  "source": "robotwin_generated",
  "hdf5_source": "/path/to/episode0.hdf5",
  "resolution": "640x480",
  "frame_count": 121,
  "action_dim": 14,
  "instructions": {
    "instruction":   "In a fixed robotic workspace, generate a rigid... [seen template, correct]",
    "instruction_1": "In a fixed robotic workspace, generate a rigid... [seen template, variant]",
    "instruction_2": "In a fixed robotic workspace, generate a rigid... [unseen template]"
  }
}

The three instructions match the WorldArena official evaluation format. instruction_2 uses unseen object descriptions for generalization testing.

---

camera_params/head_intrinsic_params.json — Camera Intrinsics

Sensor: RoboTwin Large_D435 (fixed head camera, identical for all 509 episodes and all 121 frames)

{
  "intrinsic": {
    "fx": 717.284,
    "fy": 717.284,
    "ppx": 320.0,
    "ppy": 240.0,
    "width": 640,
    "height": 480
  }
}

Parsed by load_camera_params_from_json() in diffsynth/utils/action_utils.py. Compatible with both crop and stretch resize modes.

---

camera_params/head_extrinsic_params.json — Camera Extrinsics

Convention: camera-to-world (c2w), 4 × 4 homogeneous matrix.
Note: The training code computes w2c = inv(c2w) internally — do not invert before loading.

[
  {
    "extrinsic": {
      "rotation_matrix": [
        [1.0,  0.0,   0.0],
        [0.0, -0.8,   0.6],
        [0.0, -0.6,  -0.8]
      ],
      "translation_vector": [-0.032, -0.450, 1.350]
    }
  },
  ... (121 identical entries — camera does not move)
]

Camera position in world frame: (-0.032, -0.450, 1.350) meters.
All 121 entries are identical (fixed camera). The list format is kept for API compatibility with per-frame extrinsics.

---

metadata_train_a2v.jsonl — Training Index

One JSON per line, 509 total. This is the primary entry point for ABot A2V training.

{
  "video": "episodes/beat_block_hammer__episode0/video.mp4",
  "prompt": "In a fixed robotic workspace, generate a rigid, physically consistent embodied robotic arm. The arm maintains high stability with no deformation and enters the frame to Strike the block downward with the hammer head.",
  "action_path": "/abs/path/to/episodes/beat_block_hammer__episode0/actions_16d.npy",
  "intrinsic_path": "/abs/path/to/camera_params/head_intrinsic_params.json",
  "extrinsic_path": "/abs/path/to/camera_params/head_extrinsic_params.json",
  "original_size": [480, 640]
}

Field notes:

• video: relative path (resolved by ToAbsolutePath(DATASET_BASE_PATH))
• action_path, intrinsic_path, extrinsic_path: absolute paths on the source machine — update to your local paths before training (see Quick Start)
• original_size: [H, W] = [480, 640]
• prompt: uses instruction (seen template, correct description)

---

Task Coverage

509 episodes across all 50 WorldArena Challenge tasks:

Range	Tasks	Episodes per task
All 50 tasks	Full coverage	~10 per task
High-frequency tasks	place_cans_plasticbox, stack_bowls_two, blocks_ranking_rgb, ...	10
Low-frequency tasks	beat_block_hammer, click_alarmclock, ...	10

Task distribution is uniform (10 episodes each, except click_bell which has 19 due to early collection).
The WorldArena official test set has non-uniform distribution (place_cans_plasticbox: 8.4%, ...). If you want to weight training samples by test frequency, refer to task_distribution.json in the challenge repository.

---

Quick Start: ABot-PhysWorld A2V Training

1. Clone and install paths

After downloading the dataset, update the absolute paths in metadata_train_a2v.jsonl:

DATASET_PATH=/your/local/path/to/training_data

# Replace all absolute paths in JSONL
sed -i "s|/mnt/afs/davidwang/workspace/WorldArena_Challenge2026/training_data|${DATASET_PATH}|g" \
    ${DATASET_PATH}/metadata_train_a2v.jsonl

2. Verify data integrity

import json, numpy as np
from pathlib import Path

base = Path("/your/local/path/to/training_data")
with open(base / "metadata_train_a2v.jsonl") as f:
    records = [json.loads(l) for l in f]

print(f"Total records: {len(records)}")  # 509
for rec in records[:3]:
    a = np.load(rec["action_path"])
    assert a.shape == (121, 16), f"Bad action shape: {a.shape}"
    assert Path(rec["intrinsic_path"]).exists()
    assert Path(rec["extrinsic_path"]).exists()
print("All checks passed")

3. Launch A2V training (1× H100, VACE only)

cd /path/to/ABot-PhysWorld/training

DATASET_BASE_PATH=/your/local/path/to/training_data \
DATASET_METADATA_PATH=/your/local/path/to/training_data/metadata_train_a2v.jsonl \
HEIGHT=480 \
WIDTH=640 \
CHUNK_NUM_FRAMES=121 \
NUM_GPUS=1 \
NUM_EPOCHS=3 \
LEARNING_RATE=5e-6 \
OUTPUT_PATH=/your/output/checkpoints/a2v_v1 \
bash run_train_a2v.sh

Required environment variable (optional but recommended):

# If you have an SFT DiT checkpoint, provide it for better initialization:
DIT_CHECKPOINT=/path/to/sft_dit.safetensors \

Without DIT_CHECKPOINT, the script will use base Wan2.1-I2V-14B-480P weights (still works, slightly lower quality).

---

Known Issues & Notes

1. Absolute paths in JSONL: action_path, intrinsic_path, extrinsic_path are absolute paths from the source machine. Run the sed command above to localize them.

2. Official episodes lack actions_16d.npy: The official_episodes/ split only provides actions.npy (joint-space, 14-dim). For inference/D5, you need to extract endpose from the original official HDF5 files (which contain endpose/left_endpose and endpose/right_endpose) to generate actions_16d.npy.

3. Frame count: Raw simulation produces 288 frames; actions_16d.npy and video.mp4 are uniformly downsampled to 121 frames (matching WorldArena evaluation requirement).

4. Gripper units: actions_16d.npy[:, 7] and [:, 15] are in [0, 120] (0 = closed, 120 = open), pre-scaled for get_traj_maps() colormap normalization. actions.npy[:, 6] and [:, 13] are binary {0, 1}.

5. Camera extrinsics: All 121 entries in head_extrinsic_params.json are identical (fixed-mount camera). The list format exists for compatibility with datasets that have moving cameras.

---

Data Generation Pipeline

RoboTwin 2.0 Simulation (SAPIEN 3.0 + Vulkan ray tracing)
  ├── Camera: Large_D435, fixed mount at (-0.032, -0.450, 1.350) m
  ├── Resolution: 640×480 @ ~6 fps (162 s/episode)
  ├── Output: episode{N}.hdf5 (865 MB each, includes depth)
  └── Post-processing (scripts/04_convert_to_training.py):
        ├── Uniform temporal sampling: 288 → 121 frames
        ├── Video: ffmpeg H.264 re-encode
        ├── Actions: HDF5 endpose → 16-dim + ×120 gripper scaling
        └── Instructions: RoboTwin template → official WorldArena prefix format

scripts/05_gen_actions16d_and_camera.py  →  actions_16d.npy + camera_params/
scripts/06_gen_training_jsonl.py         →  metadata_train_a2v.jsonl

---

Citation / Contact

This dataset is produced for the CVPR 2026 WorldArena Challenge Track 1.
For questions, contact: davidwangxutao@gmail.com