Skip to content

Latest commit

 

History

History
110 lines (88 loc) · 6.27 KB

action-demonstrations.md

File metadata and controls

110 lines (88 loc) · 6.27 KB

Action demonstrations: preserve the motor evidence

The dvidia.action-demonstration/v1 format records aligned observations and explicit actuator targets for the PiPER simulation profile. Validation and inspection use Python's standard library. They do not run a simulator, move a robot, train a model, recover actions from video, or qualify a skill.

The first example, authored-example.json, is a synthetic metadata fixture. It is not a corrective demonstration or a camera dataset, and contains no real pixels. Strict target export refuses authored fixtures by default. Passing a format check does not turn this example into evidence of successful manipulation.

Preserve the actual PiPER coordinates

PiPER has six radian arm targets and one native finger-slide target in metres: zero means closed and 0.035 means open. There are seven native control channels. The two measured finger coordinates are coupled signed positions; the command is not a measured full gripper gap.

The older episode tape describes a different authored arm and two jaw controls. Its limits, eight native controls and full-gap convention cannot be relabeled as PiPER data. Existing PiPER observations also lack camera frames and jaw velocity. This recorder does not automatically convert them, infer missing actions from images, or fill absent measurements with zeros.

Record one causal control boundary

Create Recorder(header, initial, initial_applied_targets), then call append(step) for each recorded interval. Each step retains:

  1. The before observation available to the decision.
  2. The requested action, separately from the issued command after any declared correction, limiter or override.
  3. The interval's applied final native seven-channel target.
  4. The after measurement, which belongs to the response rather than the current decision's inputs.

Applied targets are commands, not measured joint positions or contact forces. The final native target does not describe every intermediate 1 ms command or certify contact peaks within a control interval. initial_applied_targets means the producer's starting ramped position reference, not measured joint positions or the compensated native control. Declared ranges may tighten the packaged PiPER ranges; rehashing cannot widen them. Endpoint ramp changes must respect declared rate bounds, which still cannot certify intermediate ticks.

Finish with an explicit complete or interrupted status, task result, reason, and applicable scorer/evidence hashes. Success remains separate from recording completeness. Missing, stale and partial evidence stays explicit; incomplete records are not silently repaired into training pairs. The recorder is in memory. Save an interrupted prefix explicitly if the producer stops; this version is not a crash-safe capture service.

Bind time, geometry and provenance

Headers bind the supported profile, scene, channel order and limits, together with calibration and source identities. Keep related source recordings, sessions, corrections and retries together when assigning dataset splits; adjacent frames are not independent demonstrations.

Only the declared simulation clock is supported. Device and hardware clocks are rejected. Signal availability must precede its decision; camera exposure must also end before that decision. A frame captured earlier but received afterward cannot become a causal input. Missing readings have null values and times. Acquisition and frame identities remain ordered across missing gaps; observations are never forward filled. Native timestamps are preserved with 1 ns comparison tolerance, including roundoff at the 20 s cap. Actual longer recordings and intervals beyond the 20 ms tolerance are refused.

Camera references bind staged frame artifacts by hash. Optional local artifact readback checks those bytes, not depth accuracy, tracker quality or physical calibration. Profile, scene and calibration hashes establish internal consistency; they cannot authenticate a real recording or validate its physics.

Inspect before exporting targets

From a checkout with the package installed:

python -m dvidia_training.action_demonstration validate examples/action-demonstration-v01/authored-example.json
python -m dvidia_training.action_demonstration inspect examples/action-demonstration-v01/authored-example.json
python -m dvidia_training.action_demonstration targets --help
python -m dvidia_training.action_demonstration replay-check --help

validate_demonstration(document, require_targets=False) checks the data contract. load_demonstration reads it; write_demonstration creates an exclusive new file. inspect_demonstration summarizes it without changing it.

learning_pairs applies stricter completeness and fresh required-signal/camera checks, then requires local frame-byte readback; CLI targets uses --artifacts. Its standalone JSON retains profile/binding/calibration context, source kind and outcome as non-feature metadata, plus ordered residual units: six radians, then one finger-slide metre. Inputs are named explicitly; current issued targets, applied controls and future state remain excluded. Export preserves complete failure and unverified episodes as motor labels. It does not select good corrections: the proposed experiment must separately review successful operator segments and their scoring evidence before fitting. Source declarations and hashes cannot establish that these are good actions.

replay_check(reference, replayed) compares two recorded demonstrations. It does not perform a physics replay, execute targets or establish task competence. Its strict checks refuse missing, stale or partial evidence. This alignment tool prepares a future action-learning experiment; no new policy is trained by recording, exporting or comparing these files.

Installed Terminal command

DVIDIA Training 0.1.2 adds dvidia-action-demo, with the same validate, inspect, targets and replay-check commands as the module CLI. Validation/inspection are standard-library data operations; no GPU, simulator, network or robot connection is used. The base package installation still has its documented NumPy dependency. Real synchronized capture and reviewed corrective demonstrations remain separate work.