Spaces:
Running on CPU Upgrade
Running on CPU Upgrade
Commit History
-
Support IMU-less cameras, dcam_* filenames and camera provenance
Three upstream changes the Space needs before it can process Orbbec Gemini 305 recordings. 1. IMU-less episodes. The 305 has no IMU, and four files assumed every episode carries one. offline_vslam.cpp dereferenced calib["imu_to_cam"] unguarded and aborted with nlohmann type_error.302; convert.py treated the IMU json as a required input; checks/recording.py errored on its absence and required imu_to_cam in the calib; check_dataset.py anchored discovery on it, making 305 episodes invisible. All relaxed, without weakening the OAK-D case: a present-but-empty IMU file still errors. 2. Episode files renamed oakd_* -> dcam_*, since a Gemini recording was writing filenames claiming to be an OAK-D's. Readers resolve through the new episode_files module, which prefers the canonical name and falls back to the legacy one; episodes already on the Hub are NOT rewritten and keep working. 3. metadata.json now carries a depth_camera block naming the model, serial and fi…590a794
-
Remove legacy pre-reorg duplicates (flat scripts + episode_check/visualize)
ClaireHzl +1: Claude Opus 4.8 (1M context)Dead code superseded by the nested scripts/ layout and checks/recording; unreferenced by the API chain, CLI, and Dockerfile. Matches the monorepo. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>8b36886 -
Merge branch 'dataset-filtering-preview' into filtering-plus-tags
ClaireHzlbd07ae9
-
Port dataset-filtering changes from grabette monorepo (is_lost masking)
ClaireHzl +1: Claude Opus 4.8 (1M context)Vendors the packages/grabette-postprocess changes from the grabette monorepo branch 60-dataset-filtering-for-training (base 2a496b6): - trajectory.trajectory_to_poses: keep the SLAM's held pose on lost frames instead of zeroing them — zeroing teleports the pose to the origin, turning every tracked<->lost transition into a physically-impossible jump in the derived delta actions (the main "SLAM glitch" seen downstream). Also returns early on an empty trajectory. - dataset.build_dataset: add an additive `is_lost` (float32, per-frame) feature so lost frames can be MASKED at training time. Not prefixed observation./ action., so lerobot's dataset_to_policy_features skips it (verified: excluded from policy I/O) — the policy never sees it as input/output. - generate_dataset.py: cosmetic docstring placeholder. Isolated branch from origin/main — none of the checks/sync/tags work. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>41761e4 -
Relax angle check, make sync checks opt-in, slim content check
ClaireHzl +1: Claude Opus 4.8 (1M context)- episode content check (checks/recording.py): reduce to presence + non-empty of the SLAM inputs, angle_data.json and raw_video.mp4; validate videos from the header (decode at most one frame) instead of full-stream counting; drop the frame-drop/count/staleness heuristics. - angle check: warn only when the distal or proximal joint is genuinely static (whole-episode peak-to-peak range < ANGLE_STATIC_RANGE_RAD ≈ 0.57°), replacing the 60%-duplicate-pairs heuristic. - sync checks: now opt-in (run_sync, default off) — added a "Run sync checks (debug)" checkbox to the Space and threaded run_sync through pipeline (precheck_episodes/run_slam/build_lerobot) + the REST API. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>74f1250
-
Update grabette-postprocess (new check_sync)
Verified signaturedee00d5
-
Upload folder using huggingface_hub
Verified signaturecffabad
-
Upload folder using huggingface_hub
Verified signature261137b
-
Upload folder using huggingface_hub
Verified signature6613895 -
Upload folder using huggingface_hub
Verified signature4b921b7
-
Upload folder using huggingface_hub
Verified signature4dc78af
-
Upload folder using huggingface_hub
Verified signature0ed3076