| license: mit | |
| language: | |
| - en | |
| tags: | |
| - robotics | |
| - simulation | |
| - demonstrations | |
| - research | |
| # DVIDIA PiPER reference training data | |
| This card describes the derived training inputs for both retained reference-model candidates. This data directory contains an identity receipt; the benchmark does not redistribute raw videos, caches or reference arrays. Both candidate model files are included separately. | |
| The data comprises 96,000 control rows from 96 existing simulation episodes in 32 training starting-position groups: 64 authored-teacher demonstrations covering both destinations, followed by 32 previously collected mixed-control corrections. These are synchronized simulated camera/action data, not recovered actions from human video. Development and confirmation episodes are excluded from fitting. | |
| Each parent row contains 269 causal features, the issued-teacher or separately recorded teacher-query target, group ownership and provenance. Each reference export adds two C-order float64 `(N, 7)` arrays: exact previous normal servo references and feasible target endpoints. Six dimensions use radians; the seventh is the single-finger jaw coordinate in metres. Reset references come from the recording reset; later references come from the preceding step's actual normal ramp state. The original 20 updates are applied to teacher labels. For mixed control this is counterfactual normal-reference supervision, not a statement that those labels were executed. | |
| The parent cache was built through the original strict exporter. The reference-export construction verified sealed recording metadata and caller-pinned parent data; cache/reference loads validate derived bytes without re-reading raw RGB-D frames. This integration experiment did not revalidate raw recordings. Array, protocol, stage, source, parent dataset and manifest identities are in [training-input-identity.json](../evidence/training-input-identity.json). | |
| Source: [DVIDIA training at the benchmark commit](https://github.com/Dvidia-Inference/dvidia-training/tree/78ea4fbb4b2733e7738eb425eeecd7cb67a52195). Training uses only training-data normalization. Current teacher actions, future observations and native object/contact truth are excluded from features. Missing visual observations remain masked, and historical estimates retain their age semantics. | |
| These bounded synthetic demonstrations do not establish transfer to shoes, arbitrary videos, other arms or physical hardware. Model training losses and a single training-trajectory replay do not establish generalization. The original dataset ownership and failure accounting are preserved. | |
| DVIDIA-generated derived data and accompanying documentation use [MIT](../LICENSE). Upstream simulator and robot asset licenses remain separate. No raw human footage, credentials or upstream model weights are included in this benchmark. | |