chomeed/mimicgen_threading_d0_224x224_seg

This dataset was created using LeRobot.

Dataset Description

Every episode of chomeed/mimicgen_threading_d0_224x224 cut down to a single grasp → terminal clip. Nothing else changes: same task, same cameras, same 20 fps, same feature schema. Approach and post-outcome frames are gone, so what remains is the part of the episode where the needle is in hand and the insertion is being attempted.

170509 of 224508 source frames survive (76%), across 1000 of 1000 episodes.

Termination function

An episode is cut to the closed interval [grasp, terminal].

Start — grasp. The first frame at which the end-effector is within 80 mm of the needle and stays there: more than 90% of the next 40 frames are also under the threshold, so a fly-through on the way to the grasp does not open the clip. The end-effector and needle body frames have offset origins, so this distance bottoms out near 59 mm and never reaches zero — a 50 mm threshold fires on only a handful of episodes, 80 mm fires on all of them. Measured on observation.state[0:3] (eef) against observation.state[9:12] (needle).

End — terminal. Two cases, and which one applies is recorded per episode as rewarded:

1. The episode is rewarded → the terminal is its first frame with reward > 0, the moment the needle is judged threaded. Anything after it is post-outcome.

2. The episode is never rewarded (every rollout in the _failure split, plus a handful of demos whose reward channel never fires) → the terminal is the frame, at or after the grasp, whose needle→tripod offset is closest in Euclidean distance to the offset a successful episode has at its own terminal:

   terminal = argmin over t >= grasp of  || (needle[t] - tripod[t]) - ref ||
   ref = [0.8, 31.0, 89.4] mm
   

   ref is the mean of observation.state[9:12] - observation.state[23:26] at the first rewarded frame, averaged over the rollouts in chomeed/mimicgen_threading_d0_224x224_mtdit_flow_55k_success. That endpoint is a tight target state — per-axis standard deviation [6.5, 5.4, 2.7] mm — so the argmin is well posed: it picks each unrewarded episode's best imitation of the goal state, i.e. its nearest miss.

   The at-or-after-grasp restriction is load-bearing. Without it, one rollout's nearest approach precedes its own grasp and the interval comes out with negative length.

Why not the time limit. A failed rollout runs to a 400-step timeout, and in this task the robot usually goes on to knock the tripod over after missing. Cutting there would end failure clips in a scene state that no success ever visits, and would make them twice the length of a success clip. Cutting at the nearest miss instead puts failure terminals at roughly 50% of the source episode, against 97–100% for rewarded ones, and leaves both kinds of clip ending at the same event: the insertion attempt resolving.

Minimum length. A segment shorter than 20 frames (1 s) is dropped rather than shipped — it cannot carry a temporal window. No episode fell out this way.

What the trim changes

Recomputed so the shortened episode is internally consistent:

column	value
frame_index, index, timestamp	renumbered from 0 within the clip
episode_index	renumbered densely over the kept episodes
is_first	True on the new frame 0
is_last, done	True on the new last frame
is_terminal	True on the new last frame only if the episode was rewarded

is_terminal is the one that matters for RL. A rewarded clip ends in a real MDP terminal. An unrewarded clip ends at its nearest miss, which is a truncation — the episode would have continued — so it stays False and the final state stays bootstrappable.

Carried over untouched: reward, success, source_demo_index, observation.*, action.

Per-episode metadata

meta/segment_info.json records, for every source episode including the dropped ones:

{"source_episode_index": 0, "source_length": 194, "grasp_frame": 39, "terminal_frame": 180, "length": 142, "rewarded": true, "endpoint_residual_mm": 7.619936466217041, "kept": true, "episode_index": 0}

endpoint_residual_mm is how far that episode's terminal actually lands from ref. Use it to filter: for rewarded episodes it is small by construction, but an unrewarded episode that never approached the tripod at all still gets a terminal, and its residual is the honest signal that the clip is not a near miss.

Segment statistics

episodes kept	1000 of 1000
frames kept	170509 of 224508 (76%)
rewarded episodes	996 of 1000
grasp frame	mean 49, 22% through the source episode
clip length	mean 171, min 138, max 211 frames
endpoint residual	median 5.3 mm, p90 9.1 mm, max 15.3 mm

Built by experiments/0905/seg/build_seg.py; the segment rule itself is experiments/0905/seg/segments.py.

• Homepage: [More Information Needed]
• Paper: [More Information Needed]
• License: apache-2.0

Dataset Structure

meta/info.json:

{
    "codebase_version": "v3.0",
    "fps": 20,
    "features": {
        "action": {
            "dtype": "float32",
            "shape": [
                7
            ]
        },
        "observation.state": {
            "dtype": "float32",
            "shape": [
                37
            ]
        },
        "observation.sim_state": {
            "dtype": "float32",
            "shape": [
                45
            ]
        },
        "reward": {
            "dtype": "float32",
            "shape": [
                1
            ]
        },
        "done": {
            "dtype": "bool",
            "shape": [
                1
            ]
        },
        "success": {
            "dtype": "bool",
            "shape": [
                1
            ]
        },
        "is_first": {
            "dtype": "bool",
            "shape": [
                1
            ]
        },
        "is_last": {
            "dtype": "bool",
            "shape": [
                1
            ]
        },
        "is_terminal": {
            "dtype": "bool",
            "shape": [
                1
            ]
        },
        "source_demo_index": {
            "dtype": "int64",
            "shape": [
                1
            ]
        },
        "observation.images.agentview": {
            "dtype": "video",
            "shape": [
                224,
                224,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 224,
                "video.width": 224,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.fps": 20,
                "video.channels": 3,
                "has_audio": false,
                "video.g": 2,
                "video.crf": 30,
                "video.preset": 12,
                "video.fast_decode": 0,
                "video.video_backend": "pyav",
                "video.extra_options": {},
                "is_depth_map": false
            }
        },
        "observation.images.eye_in_hand": {
            "dtype": "video",
            "shape": [
                224,
                224,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 224,
                "video.width": 224,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.fps": 20,
                "video.channels": 3,
                "has_audio": false,
                "video.g": 2,
                "video.crf": 30,
                "video.preset": 12,
                "video.fast_decode": 0,
                "video.video_backend": "pyav",
                "video.extra_options": {},
                "is_depth_map": false
            }
        },
        "observation.object": {
            "dtype": "float32",
            "shape": [
                28
            ]
        },
        "observation.robot0_eef_pos": {
            "dtype": "float32",
            "shape": [
                3
            ]
        },
        "observation.robot0_eef_quat": {
            "dtype": "float32",
            "shape": [
                4
            ]
        },
        "observation.robot0_eef_vel_ang": {
            "dtype": "float32",
            "shape": [
                3
            ]
        },
        "observation.robot0_eef_vel_lin": {
            "dtype": "float32",
            "shape": [
                3
            ]
        },
        "observation.robot0_gripper_qpos": {
            "dtype": "float32",
            "shape": [
                2
            ]
        },
        "observation.robot0_gripper_qvel": {
            "dtype": "float32",
            "shape": [
                2
            ]
        },
        "observation.robot0_joint_pos": {
            "dtype": "float32",
            "shape": [
                7
            ]
        },
        "observation.robot0_joint_pos_cos": {
            "dtype": "float32",
            "shape": [
                7
            ]
        },
        "observation.robot0_joint_pos_sin": {
            "dtype": "float32",
            "shape": [
                7
            ]
        },
        "observation.robot0_joint_vel": {
            "dtype": "float32",
            "shape": [
                7
            ]
        },
        "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
        }
    },
    "total_episodes": 1000,
    "total_frames": 170509,
    "total_tasks": 1,
    "chunks_size": 1000,
    "data_files_size_in_mb": 100,
    "video_files_size_in_mb": 200,
    "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",
    "robot_type": "Panda",
    "splits": {
        "train": "0:1000"
    }
}

Citation

BibTeX:

[More Information Needed]