PiPER block placement v0.1
A DVIDIA research artifact for learned PiPER block placement. The selected policy is a high-resolution time-conditioned RBF feedforward motion primitive (16 cubic spatial terms ร 401 temporal centers, ฯ=0.025 s) (schema dvidia.piper-motor-primitive/v2 in model/model.json). It is trained by imitation from demonstrations generated by an authored synthetic simulator teacher, using privileged simulator data. It does not establish learning from human video, camera-only control, physical robot performance, or hardware readiness.
The selected policy metadata and weights are in model/; the shuffled-label model is retained as a negative control. The protocol and aggregate report are included. The dataset preserves all scheduled demonstrations, earlier failed MLP candidate checkpoints and their development traces, and all confirmation attempts. Browse the complete dataset or open the recorded replay viewer.
Selected v2 primitive inputs: reset-time object XY, one of the two known goal coordinates, and elapsed time. Live joint positions are used only by generic native joint-target caps. The selected policy does not consume TCP or contact estimates and does not learn contact feedback or recovery. The retained demonstrations and evaluation traces contain richer privileged simulator data; those records do not change this selected-policy input contract.
Confirmation encountered a serialization guard: the live configuration used tuples, while the persisted JSON representation used lists. A bounded adapter checked the exact error and the frozen bindings, then resumed scoring from each already-saved native episode tape; it did not rerun physics. The adapter source, root summary, and per-attempt recovery records are retained in sources/piper_imitation_driver.py, driver-recovery.json, and driver-recovery/. The published report keeps the same frozen confirmation roster and result counts, and the recovery summary records zero completed tapes reexecuted.
Run the policy in native simulation
Use Python 3.12 or later. This runs locally in pinned native simulation and does not connect to robot hardware.
python -m pip install --force-reinstall "dvidia-training[arm] @ https://github.com/Dvidia-Inference/dvidia-training/releases/download/piper-imitation-v0.1/dvidia_training-0.1.0-py3-none-any.whl"
dvidia-robot-sim fetch piper --output robot-assets
wget https://huggingface.co/Dvidia/piper-block-placement-v01/resolve/main/model/model.json -O model.json
dvidia-piper-learned --model model.json --assets robot-assets --start-x 0.30 --start-y -0.03 --goal 0 --output learned-run
The feedforward primitive does not learn reactive contact correction or recovery; the native servos remain authored.
Original DVIDIA artifact files are released under the MIT License. The PiPER model and other upstream components retain their own licenses.
The report declares 5 variants and 7 limitation entries. Variant results are recorded below; read the report for per-attempt outcomes.
| Variant | Successes / scheduled |
|---|---|
| learned | 5 / 12 |
| teacher | 12 / 12 |
| open_jaw | 0 / 12 |
| hold | 0 / 12 |
| shuffled_labels | 0 / 12 |
Declared limitations
- Synthetic state/action demonstrations from an authored PiPER controller, not human video or teleoperation.
- Earlier MLP candidates used privileged simulator joint, TCP, object and contact measurements; no vision pipeline was trained. This broad report statement describes those candidates, not the selected v2 primitive input contract.
- Only starting-position interpolation for a 40 g rigid block and two trained destinations is evaluated.
- The pinned native servo and rate limiter remain authored low-level control; no teacher, phase machine or IK is called by the student.
- Uncalibrated simulated friction, contacts and actuators; no physical robot qualification or other-arm transfer.
- Compute latency excludes sensing, camera processing, communication and physical actuation; CPU/GPU minima are not certified.
- Simulation-only time-conditioned trajectory distillation from synthetic authored-controller demonstrations. Initial privileged object XY, two known goals and measured clock condition learned feed-forward references; native servos remain authored. No learned contact feedback, recovery, camera learning or physical robot qualification.