Skip to content

Development notes

This page keeps active, incompletely validated work separate from the package's release-facing claims. The features below may be useful for research and API review, but they are not validated simulator or robot capabilities.

Experimental feature status

Work in progress Evidence currently available Evidence still required
Skill capsules and demonstration ingestion Contract, serialization, and round-trip tests Dataset-scale curation and retrieval study
Primitive programs and recovery graphs Unit tests and a deterministic synthetic state-machine example Closed-loop simulation and real-robot task trials
Analytic and learned candidate portfolios Isolated candidate and common-gate tests Matched robot-feasibility and task-outcome evaluation
Persistent relational FATROP retargeting Solver tests and exploratory geometric fixtures Frozen confirmatory protocol with full orientation and robot constraints
Gripper and event-based segmentation Sample-exact and hysteretic change-point tests Annotated demonstrations and semantic boundary evaluation

See the primitive-program guide, provisional API inventory, and experimental validation notes for implementation details and reproducible component checks.

VLA evaluation hypothesis

Initial-frame relative pose chunks are already invariant under one shared world-frame transform. DHB-XR's research hypothesis is not that DHB is more frame-invariant. It is that a structured motion representation may provide a useful common interface for retargeting, retrieval, time normalization, tokenization, and later execution stages.

The release currently verifies numerical frame stability and software interfaces. A matched learned-policy study is still required before claiming that a DHB action head improves VLA task success, generalization, token rate, or data efficiency over an initial-frame relative-pose head.

Promotion rule

Move a development item into release highlights only when its claim is backed at the corresponding evidence level:

  1. Unit tests establish software behavior.
  2. Isolated examples establish bounded component behavior.
  3. Frozen matched benchmarks can support representation comparisons.
  4. Closed-loop simulation can support simulator-task claims.
  5. Repeated real-robot trials are required for physical-task claims.

Results at one level must not be described as evidence for a higher level.