dataset/DATASET.md

This export is sim_ep_03: clutter-only state-feedback pick/place. See TASK.json for task semantics. The upstream obstruction-task examples below describe a different dataset; the timing, camera and training interfaces apply.

UR5e obstruction-removal training data

The current run lives in runs/pi05_15hz_3variants_20261023/. It requests 171 episodes and 513 camera videos, if every generation passes. Capture, measured-state sampling and held-action commands are 15 Hz; physics and contact logging are 300 Hz. Each sample spans exactly 20 physics steps. The plan retains ten historical obstacle scenes for originals 0, 10, 16, 27, and 36, plus the ten scenes already generated for original 1. Each of the remaining 37 originals has three obstacle variations with surrounding clutter. Failed attempts never count as accepted episodes.

Existing 25 Hz captures remain in runs/pi05_430_three_camera_20261022/. They are not relabeled or interpolated to 15 Hz. Scene specifications and controllers are reused, followed by fresh physical execution and all three camera captures. The old camera-repair queue and separate 46-scene supplemental queue are superseded. The new pilot covers a detour, historical removal and new destination-blocking removal, including a held-out source, before scaling.

Use the repository Makefile and config/pi05_generation.json; see reproduction instructions. make start runs the pilot, full two-GPU queue, export, and actual openpi loader validation in order.

Read the current run's generation_plan.json, manifest.json, physics_status.json, render_status.json, and export reports for actual progress and accepted/failed counts. The presence of a video or an older task_success flag does not make an episode eligible for training.

Verified training interface

On 2026-10-22, the upstream openpi checkout was pinned to 215abfb217dbac7d5f1273282331b9b1866c0479. Its pyproject.toml pins LeRobot to 0cf864870cf29f4738d3ade893e6fd13fbd7cdb5, whose dataset format is v2.1. This is deliberately not an export to the latest, incompatible LeRobot layout. The exporter also requires datasets==3.6.0 from openpi's lockfile. Version 4 changes column access and Parquet feature metadata and fails the pinned loader.

References: openpi custom-data instructions, dependency pin, model/data transforms.

Use pi05_base, not the Franka/DROID expert or DROID normalization. The UR5e is a different embodiment. policies/pi05_ur5e.py defines the custom adapter and config. The model retains action dimension 32 and action horizon 50; seven real channels are padded only inside the model preprocessing.

State, actions, and time

The seven-channel state and action order is:

1. shoulder_pan_joint
2. shoulder_lift_joint
3. elbow_joint
4. wrist_1_joint
5. wrist_2_joint
6. wrist_3_joint
7. gripper_open_fraction

Arm positions are absolute radians in the imported UR5e URDF convention; velocities are radians/second. State is measured from PhysX, independently of commanded joint targets. This is a single-arm seven-channel dataset, not the historical padded 14-channel RoboPRO contract.

Gripper opening is 1 - native_position / 255 (1 open, 0 closed). Commanded opening uses the requested native actuator position. Measured opening uses the finger-joint-to-native calibration map; it is not a measured jaw aperture or a force measurement. The mapping, raw native measurement/request, finger angle, and all articulation joint positions/order are preserved in each raw synchronized.hdf5. The gripper response retains the inherited latency and rate model. Finger-drive effort is a simulator actuator setting, not calibrated physical squeezing force.

Physics runs at 300 Hz; actions and image observations run at 15 Hz. Each sample has measured state and all three camera images at simulation time t. Its absolute target is then issued and held over [t, t + 1/15). Velocity feedforward is the difference between consecutive commanded positions divided by 1/15, held throughout that control interval. Gripper command latency is part of the simulated plant. action_end_timestamp records the actual end. Rendering freezes physics, so sequential camera readbacks share the same simulation timestamp. There is no pairing of cached images with later joints. Wall-clock rendering duration is unrelated to action duration.

Cameras and model mapping

All three streams are RGB uint8, 640 × 480, with calibration in scene_spec.json:

Export feature	Source HDF5 role	openpi input
observation.images.wall	countertop_camera	base_0_rgb
observation.images.wrist	left_camera, attached to flange	left_wrist_0_rgb
observation.images.body_forward	right_camera, base-mounted	right_wrist_0_rgb

All three model image masks are true. The model's fixed third slot is named right_wrist_0_rgb; its physical camera in this dataset is the base-mounted body-forward camera, not a second wrist camera. Physical roles and calibration remain explicit in exported feature names and scene metadata. Openpi resizes with its standard 224 × 224 image transform. No arbitrary flips are applied. Camera poses use OpenCV optical axes in the scene specification. Stored flange/object poses are xyz in meters followed by quaternion wxyz. The wrist extrinsic is relative to the flange, whose measured pose is retained.

Raw PNGs remain available. LeRobot contains separate MP4 streams and Parquet numeric samples with identical frame counts. The exporter decodes every video and checks presentation timestamps against the 15 Hz sample grid.

Earlier two-camera completion method (archived 25 Hz run)

The following method applied to the archived 25 Hz run, not the new 15 Hz run. Compatible completed two-camera episodes could fill variation zero of their original recording by rendering only body_forward. The existing wall/wrist PNGs are reused unchanged. Their measured joint, action, gripper, phase, and timestamp datasets are copied bit-for-bit to a new HDF5 file; the original files remain untouched. The third view uses every saved articulation DOF (including measured gripper mimic joints) and object poses logged at the same simulation timestamp. It never interpolates states or executes new dynamics. The reconstructed flange must agree with the recorded flange within 2 mm and 0.01 rad at every frame. This is visual replay of an already validated physical execution, not a newly executed demonstration. The new camera's exposure time is the original simulation time, not its later wall-clock render time.

third_camera_status.json records progress and third_camera_completion.json records provenance, timestamp error, flange error, source checksum, and exact preservation of existing numeric datasets. A separate complete episode must pass rendering, export, and the real openpi loader before this method scales. If later episodes fail completion, the normal batch falls back to fresh capture. These reused episodes occupy existing slots in the reduced plan; they do not increase the count or displace the 50 historical obstacle variations.

October 23 camera identity correction

The initial multi-camera capture had a viewport/readback timing defect: the wrist stream often repeated the wall view. All 73 previously accepted captures were withheld after image comparison, and their previous export gates were invalidated. File counts, time alignment, and successful loader decoding alone did not establish correct camera identity.

Capture now uses a fixed render-camera prim with the requested calibrated pose and intrinsics, and at least 16 rendered frames before each readback. The requested/render-product binding and pose are checked at every exposure. camera_poses_base_opencv and camera_identity_version are recorded; acceptance checks those poses against wall/body calibration and measured flange × wrist extrinsic, and rejects near-duplicate camera views at distributed samples.

analysis.pi05_camera_recovery first repairs episode 31 and exports it through the actual openpi loader. It then repairs the remaining completed captures. Reference images are rendered at evenly spaced frames and every task-phase transition to determine which streams need replacement. Correct streams are retained, while faulty streams are rendered from the original physical states. All existing robot numeric datasets must remain bit-for-bit equal. Corrected training captures live in camera_repaired_v1/; prior videos/ addresses point to the corrected videos once their replacement passes, with old videos saved under videos_before_camera_repair/. Original raw captures remain for audit. camera_repair_status.json reports actual repaired counts. These repairs belong to the archived 25 Hz run. The current 15 Hz pipeline uses the corrected capture implementation and has its own pilot gate.

Parallel execution on both GPUs

The current coordinator is analysis.pi05_parallel: two persistent Isaac processes, one per RTX 4080 SUPER. Each process uses explicit CUDA-index renderer selection (active_cuda_gpus), the matching physics GPU setting, a separate portable Kit directory, and a GPU-specific lock. Workers render offscreen; GPU 1 cannot present a window through the desktop display attached to GPU 0. The GPU 1 offscreen preview passed the same camera-identity comparisons.

Each worker owns a different episode/variation output directory and writes an isolated result file. Only the coordinator updates aggregate progress and records accepted videos. It schedules fresh physical generation and three-camera capture, followed by unified export and openpi validation. Existing 25 Hz captures remain separate. Physical failures remain reported; parallel execution does not relabel them successful. Two workers do not imply twice the measured throughput: CPU physics, PNG writes, and encoding can leave a GPU briefly idle.

Monitor parallel/status.json, parallel/gpu0/current.json, parallel/gpu1/current.json, and the per-GPU worker logs. Each completed job records elapsed time in parallel/results/. The coordinator can be launched with make start when no coordinator/workers are running. Use make status and make estimate for progress and an updated timing estimate.

Splitting and normalization

Validation original recording IDs are 7, 14, 21, 28, 35; the remaining 38 IDs are training sources. All attempts and future augmentations of a recording inherit that assignment. Failed sources are reported, never silently replaced by a source from the other split.

Training and validation are separate local repositories: local/ur5e_obstacles_15hz_train and local/ur5e_obstacles_15hz_validation. This prevents openpi's default all-episode LeRobot loader from accidentally training on held-out recordings. Set HF_LEROBOT_HOME to the export root.

Stored actions remain absolute. The adapter converts the six arm channels in each 50-step (3.333-second) action chunk to deltas relative to the first measured state in that chunk. The gripper remains absolute. At episode ends, actions repeat the final command according to LeRobot's padding convention; chunks never cross episode boundaries. Openpi's action padding mask is available from LeRobot; the standard training transform uses endpoint repetition.

assets/ur5e/norm_stats.json is calculated only from accepted training sources, after that same action conversion, including chunk endpoint padding. It stores mean, standard deviation, and exact 1st/99th percentiles. Pi0.5 uses quantile normalization; validation reuses training statistics. The loader smoke test checks the normalization/delta inverse against stored commands.

Acceptance and provenance

For newly sampled removal tasks, the obstruction occupies the original destination measured from the real final image. The controller must physically pick it, move it to a clear table location, release it, then grasp and place the original target. During physical execution objects stay dynamic, with no attachment, teleportation, or object-pose playback. Adding a missing camera later uses the saved physical trajectory as described above, while retaining its original validation. Planner revision 2 centers the obstruction between nearby rack pegs while checking that it still overlaps the observed destination. It can shift the target's planned placement by at most 13 mm toward that gap; acceptance always compares against the unchanged observed destination with the original 15 mm tolerance. This avoids spawning a blocking cup tipped over a peg. Surrounding props use existing textured meshes with explicit mass, friction, scale, and source metadata. Layout seeds and rejected attempts are retained.

Acceptance requires target grasp/lift, correct rack placement within 15 mm in horizontal position and height, sustained rack support and release, and terminal speed below 0.02 m/s. The obstruction must have a two-finger grasp, rise over 70 mm, finish within 40 mm of its clearing point, be released, and settle. Unplanned robot-object/table/rack contacts and obstruction-clutter contacts above 0.05 N are rejected. Surrounding bodies must move less than 10 mm. Contact evidence is sampled at the 300 Hz physics rate. These checks cover reported PhysX actor contacts; they are not a hardware safety proof. The inherited robot articulation disables self-collision. These checks therefore cover robot contact with objects, rack, and table, plus unintended prop contacts; they do not certify clearance between the robot's own links.

training_validation.json adds numeric finiteness, joint ordering, initial robot state, action timing, gripper convention, camera synchronization, and image-count checks. Raw HDF5, full contact traces, summaries, source identities, instructions, specs, and failure files remain alongside each generation. The original tar archive is read-only throughout; its SHA-256 is retained.

Reconstruction uses calibrated RGB silhouettes and wrist-depth table fits. Overlay reviews and fit quality are retained per original recording. Mass, friction, gripper response, and rigid plastic geometry remain simulator priors; the photographic table/background projections do not constitute a full room reconstruction. This dataset is simulation training data, not validated real-robot transfer data.

Running and checking

The current batch script refuses to scale until pilot_export/openpi_loader_validation.json records successful three-camera end-to-end pilots covering new removal, historical removal, and historical detour. Run simulator commands under the shared /tmp/real2sim_isaac.lock when using the old serial runner. The parallel coordinator instead uses /tmp/real2sim_isaac_gpu0.lock and /tmp/real2sim_isaac_gpu1.lock; never run the serial and parallel runners together.

make doctor
make start
make status
make estimate

Use the pinned openpi and LeRobot source directories on PYTHONPATH for export and analysis.pi05_openpi. The isolated data-validation environment is under the run's deps/venv; its exact package versions are recorded with the run. openpi_loader_validation.json establishes actual video decoding, action chunking, tokenizer/preprocessing, normalization inversion, and minibatch construction. It does not claim that an optimizer step or policy fine-tuning has been performed. No automatic upload is part of this workflow.

The local TorchCodec runtime additionally uses the Ubuntu FFmpeg 4 shared libraries extracted into deps/ffmpeg4/usr/lib/x86_64-linux-gnu. Include that directory in LD_LIBRARY_PATH; no system packages were changed. The durable orchestrator supplies this and the pinned source imports automatically:

make run

Generation stages resume using their per-source/per-variation status files. Each accepted render writes three viewing MP4s under its videos/ directory alongside raw synchronized HDF5 and PNGs. These viewing files precede the final LeRobot export; training readiness requires the export and loader checks. The final exporter preserves original recording and variation IDs. Failed attempts are not counted as successful variations or silently filled by duplicates. Export requires a new output directory; preserve an interrupted export before restarting that stage.

Reusing the historical obstacle videos

analysis.pi05_legacy_audit audits the canonical 50 obstacle entries in the old video inventory, verifies their video checksums, and extracts logged joints to legacy_obstacle_reuse/episodeXXXX/obstacle_YY/logged_joints.hdf5. These files preserve 50 Hz measured arm states, commanded targets, calibrated gripper state, camera exposure times, and the old MP4 resampling map. Their seven-channel order matches this contract, but they are explicitly not training-ready episodes. Old camera captures were typically 72 ms apart; some have no joint measurement within 8 ms. Old videos repeat captures on a 15 fps presentation grid and can display an image up to 72 ms old. They also lack the wrist stream. No interpolated state is relabeled as an exact measured observation.

The audit found 46 eligible scene/trajectory candidates: 29 clearing and 17 detour variants. Four old runs fail the stricter screen (two move the rack more than 10 mm and two have unintended robot contacts). The old artifacts are preserved. Reuse means new physical execution and synchronous 15 Hz capture from the saved scenes/plans, with fresh acceptance checks on every rollout.

The former analysis.pi05_legacy_reuse supplemental queue is cancelled. The current settings-driven coordinator includes all 50 source scenes, adds surrounding clutter while retaining their obstacle/task, and retries fresh physical execution. Its pilot covers clearing and detour tasks through all three cameras, export, and openpi loading before scaling. It records failures rather than treating old acceptance as proof of new success. Detours retain avoidance instructions and the legacy_detour profile; old source/route-clearing demonstrations use legacy_clearance. Neither supplemental profile claims that the original destination was occupied, and neither relaxes the main required_clearance profile. Both retain the new timing, grasp/release, settling, destination, and unintended-contact requirements as applicable to the labeled task.

The unified export uses local/ur5e_obstacles_15hz_{train,validation} and computes normalization using only its training rows. Original IDs 0, 10, 16, 27, 36 are all training sources under the fixed split; their augmentations never enter validation. The adapter reads the dataset name from export_report.json. Monitor the current run's per-variant physics_status.json, render_status.json, failed_generations.json, and manifest.json for actual results; preparation is not acceptance.