Jev Robot

No. It runs Mapika/decider-2b, a community open decision model, behind a Jev-style typed-choice interface, and describes itself as inspired by Jev-style decision methods and a PiPER demonstration. No hosted API key is required; the pinned roughly 3.76 GB checkpoint is downloaded separately.

Deterministic execution checks can reject a choice, but they never substitute another skill. Object motion or unexpected contact interrupts execution and triggers another model decision from fresh sensor-derived facts, and physics keeps running while the next inference is pending.

The frozen controller passed 10 of 10 runs on previously unused seeds 20 to 29 with randomized initial positions, and completed 30 of 30 object relocations. An earlier development batch passed 8 of 10, and its failures and fixes are preserved alongside the successful runs rather than discarded.

Across 141 genuine model calls on an RTX 5090 Laptop GPU, median decision latency was 32.6 ms with 34.3 ms at P95, and every call supplied the same ten candidates and returned zero generated text tokens. The published numbers use PyTorch 2.10, CUDA 13.0 and Triton 3.6, and the CUDA build has to match your driver and GPU.

Sensor-derived geometry, visibility, contact and execution history, described in English by a hand-written adapter. The model gets no images and no simulator object poses. It returns one of ten fixed skills with its confidence and the complete probability distribution, using a forward pass over decision labels rather than generating control code.

Python 3.13 for simulation and 3.11 for inference, with ffmpeg and libx264 for recording and EGL or OpenGL for rendered perception; on Ubuntu install ffmpeg, libegl1, libgl1, libgl1-mesa-dri and libxcb1 first. Serve the model with scripts/serve_decider.py, then run piper_sim with --policy decider, since the CLI defaults to the scripted baseline. Scope is one red cube, a fixed tray and an articulated PiPER model, with no physical robot deployment. Project code is Apache-2.0 and the bundled PiPER assets stay MIT.

The output directory holds an MP4, an English HTML report you open at outputs/dynamic/index.html, plus observations, decisions, disturbances, trajectories and physical validation metrics. Exit status is 0 only when all required relocations and physical success checks pass, so a partial run is distinguishable from a clean one.