README.md

Generate RGB Data

This step converts state-based trajectories into two-view RGB observations (third-person view + hand view) using ManiSkill3.

Prerequisites

• The .h5 trajectory file and its corresponding .json metadata file must be placed in the same directory.
• The provided dataset contains 1,100 state-based trajectories for the StackCube task.

Usage

Run the following command within your ManiSkill3 environment:

python -m mani_skill.trajectory.replay_trajectory \
    --traj-path /path/to/stackcube_000_0_1100.h5 \
    --use-first-env-state -c pd_ee_delta_pose -o rgb \
    --save-traj --num-envs 10 -b physx_cpu

Adjusting Image Resolution

The default resolution is 128×128. To change it, edit the .json metadata file and add the width and height fields under sensor_configs. For example, to set the resolution to 256×256:

{
    "env_info": {
        "env_id": "StackCube-v1",
        "env_kwargs": {
            "obs_mode": "none",
            "control_mode": "pd_joint_pos",
            "render_mode": "rgb_array",
            "sensor_configs": {
                "shader_pack": "default",
                "width": 256,   // ← add this
                "height": 256   // ← add this
            },
            "human_render_camera_configs": {
                "shader_pack": "default"
            },
            "viewer_camera_configs": {
                "shader_pack": "default"
            },
            "sim_backend": "auto"
        },
        "max_episode_steps": 50
    }
    // ...
}