| license: mit | |
| language: | |
| - en | |
| tags: | |
| - robotics | |
| - simulation | |
| - openusd | |
| - mujoco | |
| - agilex-piper | |
| - dvidia | |
| pretty_name: DVIDIA task environments v0.2 | |
| size_categories: | |
| - n<1K | |
| # DVIDIA task environments v0.2 | |
| Reusable local environment recipes for a pinned native PiPER arm, a five-collider open bin, and up to four original rigid object instances. A rigid shoe surrogate is compound box geometry; it is not a measured or deformable shoe. All dynamics parameters are authored assumptions. | |
| The frozen passive benchmark retains 6 positive native scenes, 6 negative completion fixtures, and 4 invalid-layout prechecks. All 12 native outcomes matched their declarations; all 4 invalid layouts were rejected. A separate read-only audit reconstructed 13,262 recorded 1 ms state/goal samples and 1,337 saved replay frames with zero consistency mismatches. These counts measure scene and verifier checks, not a learned robot manipulation success rate. | |
| [Open the environment studio and recorded benchmark](https://dvidia-robot-simulation-lab.static.hf.space/task-environments/v02/index.html) · [Source and Terminal tools](https://github.com/Dvidia-Inference/dvidia-training) | |
| Download `task-environments-v02.zip`, extract it, and serve the directory with `python -m http.server 8000`. The browser can edit a draft and replay saved native motion. Running new physics requires the local Python runtime and verified PiPER assets; a static hosted page does not itself run the native simulator. See `guide.md` for installation, local UI/Terminal commands, and supported inputs. | |
| PiPER holds its native home controls. No grasp policy is trained, no video-to-skill learning or physical-robot transfer is demonstrated, and the consistency audit does not establish real-world physics or sensor accuracy. Negative controls explicitly disable gravity in two stationary verifier fixtures and set initial velocity with zero friction in one moving fixture; all interventions and attempts are retained. Geometry is stored once by content hash. | |
| Original DVIDIA code and geometry: MIT; pinned PiPER simulation assets: MIT with notices in the archive. Other dependencies retain their own licenses. DVIDIA research · [@iammrriver](https://x.com/iammrriver) · hello@dvidia.org. | |