Miical/pi05-libero10-task4-recap-labeled

This dataset was created using LeRobot.

Dataset Description

This is the fixed, labeled replay dataset used for a single-iteration ReCap experiment on LIBERO-10 task 4:

Put the white mug on the left plate and put the yellow and white mug on the right plate.

The 32 trajectories were collected from Miical/pi05-libero-10-sft-step-200. The collection contains 15,064 frames at 10 FPS, with 8 successful and 24 failed trajectories. It includes the external and wrist camera videos, proprioceptive state, actions, sparse environment rewards, termination metadata, and frame-level ReCap annotations.

The dataset is intentionally small and reflects one fixed replay pool. It is published to reproduce and diagnose this ReCap experiment, not as a general LIBERO-10 training corpus. Training on these annotations did not improve held-out task performance: the source policy scored 14/50 (28%). The resulting ReCap policy scored 5/50 (10%) when evaluated with the positive ACP prompt tag; an unconditioned diagnostic evaluation scored 8/50 (16%). Continuing policy training from five to ten epochs recovered the positive-ACP result to 12/50 (24%), which remained below the source-policy baseline. Continuing to fifteen epochs reached 14/50 (28%), matching the source-policy baseline. Continuing to twenty epochs collapsed to 3/50 (6%) despite lower training loss, so the fifteen-epoch checkpoint was the best saved policy in this run.

• Source policy: Miical/pi05-libero-10-sft-step-200
• Framework: verl-vla
• Format: LeRobot v3.0
• License: apache-2.0

ReCap Annotations

• recap.return: normalized return target clipped to [-1, 0], with failure coefficient 1.0.
• recap.value: prediction from the fully fine-tuned microsoft/resnet-18 value critic using the external camera, wrist camera, and 8-dimensional robot state.
• recap.advantage: return minus the inferred value estimate.
• recap.indicator: binary ACP training label. The positive 30% are selected with future-smoothed 50-step advantages using a 50-frame window and decay 0.95.

The value critic was trained for 20 epochs on this replay pool. The annotation summary is: value mean approximately -0.537, advantage mean -0.0365, advantage standard deviation 0.0258, and positive indicator ratio 0.3001.

Dataset Structure

meta/info.json:

{
    "codebase_version": "v3.0",
    "robot_type": "panda",
    "total_episodes": 32,
    "total_frames": 15064,
    "total_tasks": 1,
    "chunks_size": 1000,
    "data_files_size_in_mb": 100,
    "video_files_size_in_mb": 200,
    "fps": 10,
    "splits": {
        "train": "0:32"
    },
    "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
    "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
    "features": {
        "observation.images.image": {
            "dtype": "video",
            "shape": [
                256,
                256,
                3
            ],
            "names": [
                "height",
                "width",
                "channel"
            ],
            "info": {
                "video.height": 256,
                "video.width": 256,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 10,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.images.wrist_image": {
            "dtype": "video",
            "shape": [
                256,
                256,
                3
            ],
            "names": [
                "height",
                "width",
                "channel"
            ],
            "info": {
                "video.height": 256,
                "video.width": 256,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 10,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.state": {
            "dtype": "float32",
            "shape": [
                8
            ],
            "names": [
                "state"
            ]
        },
        "action": {
            "dtype": "float32",
            "shape": [
                7
            ],
            "names": [
                "action"
            ]
        },
        "next.reward": {
            "dtype": "float32",
            "shape": [
                1
            ],
            "names": null
        },
        "next.terminated": {
            "dtype": "bool",
            "shape": [
                1
            ],
            "names": null
        },
        "next.truncated": {
            "dtype": "bool",
            "shape": [
                1
            ],
            "names": null
        },
        "next.success": {
            "dtype": "bool",
            "shape": [
                1
            ],
            "names": null
        },
        "info.is_intervention": {
            "dtype": "bool",
            "shape": [
                1
            ],
            "names": null
        },
        "info.reset_state_id": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "timestamp": {
            "dtype": "float32",
            "shape": [
                1
            ],
            "names": null
        },
        "frame_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "episode_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "task_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "recap.return": {
            "dtype": "float32",
            "shape": [
                1
            ],
            "names": null
        },
        "recap.value": {
            "dtype": "float32",
            "shape": [
                1
            ],
            "names": null
        },
        "recap.advantage": {
            "dtype": "float32",
            "shape": [
                1
            ],
            "names": null
        },
        "recap.indicator": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        }
    }
}

Citation

BibTeX:

[More Information Needed]