Dvidia/piper-reference-actions-v01

DVIDIA PiPER reference actions v0.1

Two compact CPU research candidates learned from 96,000 synchronized simulation demonstration rows. Both trained in about 5 seconds on the measured host and passed save/load replay checks. Neither is a qualified skill: both achieved 0/4 strict placements in the bounded native development smoke. Both candidates and negative results are retained.

This repository contains model JSON files, cards and evidence for a custom Python API. It is not a Transformers model, a qualified-skill CLI installation package or a physical robot driver. No model was selected as a winner.

Measurement	Absolute reference (Cabs)	Prior-reference delta (Cdelta)
Training rows	96,000	96,000
Fit time on measured shared host	5.230713 s	5.097628 s
Model JSON size	539,345 bytes	538,021 bytes
Saved-model replay parity	1,000/1,000 outputs identical	1,000/1,000 outputs identical
Complete native development episodes	4/4	4/4
Strict placements	0/4	0/4
Bilateral grasp contact seen	3/4	4/4
Qualified lift seen	1/4	0/4
Carry completed	0/4	0/4
Episodes with unexpected contacts	2/4	4/4
Unexpected-contact native ticks	7,606	19,751
Native callback including disk guard p95	0.295460 ms	0.284297 ms
Full offline boundary p95	22.362537 ms	22.049043 ms

Absolute model card · Delta model card · Data card · Native audit · Failure diagnosis.

Load a candidate

Install the source revision used for measurement and its pinned NumPy dependency in a virtual environment:

python -m pip install "git+https://github.com/Dvidia-Inference/dvidia-training.git@78ea4fbb4b2733e7738eb425eeecd7cb67a52195"

Download this repository's files and run Python from its root:

from dvidia_training import camera_action_reference_model as reference_model

model = reference_model.load_model("models/delta/model.json")
assert model["task_qualified"] is False
assert model["physical_robot_ready"] is False
policy = reference_model.ModelPolicy(model)
print(model["variant"], model["training"]["samples"])

This only loads and inspects the candidate; it starts no simulator or hardware. policy(context) expects the recorder's exact causal camera/encoder context, historical initial visual anchor, 160 ms history, availability masks and an actual retained float64 previous servo reference starting at the registered reset. An arbitrary image or joint vector does not satisfy that input contract. The source's new model format is not handled by the qualified-skill CLI installer.

What was measured

Exactly two real training-data fits used the same 64,000 authored-teacher simulation rows plus 32,000 earlier mixed-control correction rows. Recipe: seed 17, two 64-unit tanh layers, 50 epochs, batch 512, Adam 0.003. Parent cache and exact prior-reference/feasible-target identities are bound in the training input receipt. Raw videos and derived training arrays are not redistributed in this compact model repository.

Both candidates replayed 1,000 steps of the first training teacher trajectory. Predictions never fed into later replay observations or references; this is serialization/inference integration, not generalization. A separate native batch used the first two development groups, both goals and both unchanged candidates. All eight 20-second simulated episodes completed with no teacher intervention or fallback. No confirmation data or hardware was used.

Neither candidate completed carry or strict placement. Seven of eight episodes never reached the required sustained bilateral table-free lift. All four delta trajectories crossed the strict scorer's unchanged finite TCP coordinate envelope of ±0.5 m; this was not a NaN or solver-warning diagnosis. Cabs visual estimates were available 59.9% of decisions and Cdelta 53.7%, with longest missing intervals 13.36 s and 6.16 s. These are observations from four cases per candidate, not a causal diagnosis or a statistical ranking.

Sixteen earlier infrastructure refusals are preserved separately: eight at the wrapper's import guard, then eight at CoreGraphics initialization. A tiny static-frame preflight established scoped graphics access before the third batch. The source, models, roster, physics, guards and scorer remained unchanged across the two later batches. All-batch accounting and graphics preflight retain those boundaries. Do not pool infrastructure refusals into a 0/24 manipulation score.

Measurements used a shared macOS 26.5.2 arm64 host, Python 3.12.9, NumPy 2.5.3, MuJoCo 3.15.0 and declared BLAS threads 1. Fit timing includes validation and normalization. Whole-process integration peak RSS was 1,345,404,928 bytes, not a model-only allocation or minimum hardware requirement. Callback timing excludes camera/rendering/physics/transport; the native callback includes its disk guard. Full boundary includes file writes and alternate camera work while the simulator pauses. Its 22 ms p95 is not a 50 Hz real-time guarantee. The renderer reported a depth-accuracy limitation due to unavailable ARB_clip_control; no backend depth calibration was established.

These remain bounded PiPER block-placement research candidates. No arbitrary-video skill recovery, other-arm transfer, hardware qualification, release gate, negative-control suite or comparator superiority is claimed. The proposed 164-attempt screen remains unrun and frozen v0.1 remains 5/64 strict learned placements, unqualified. The research report and all eight recorded attempts are publicly available. Anonymous readback verified all 414 published report, diagnostic and viewer files against research commit 32fcceeb.

License and provenance

DVIDIA-generated weights, derived-data documentation and source use MIT. Dependencies and upstream robot assets retain their own licenses and are not bundled here. No raw human footage or credentials are included. Candidate files are byte-identical to the retained training outputs; their file and canonical content hashes are listed in the model cards. Runtime source was 78ea4fbb. The source, model cards and benchmark evidence were published at 8774265f; this documentation commit preserves the measured runtime above. The model's scientific_registration remains null; external receipts bind these operational experiments.