ORBIT (orbit-robotics) — RETIRED

This package is no longer maintained. Version 0.7.2 is its final release. All earlier versions have been yanked. Please do not build on it.

Why it was retired

ORBIT was sold on a claim that turned out to be false.

Versions through 0.8.0 stated that ORBIT's A–F dataset grades were "calibrated against 82 real training runs with known success rates," that ORBIT could predict whether training on your dataset would succeed, and that per-episode scoring could identify which demonstrations to keep or remove to improve a policy. Those claims are retracted.

They were retracted because we tested them ourselves and they did not hold:

• We trained 701 policies on controlled demonstration subsets — Diffusion Policy and ACT from scratch, SmolVLA and π₀.₅ fine-tuned — evaluated on PushT, two bimanual ALOHA tasks, and two ten-task LIBERO suites, every batch pre-registered.
• No score we built for a demonstration predicted whether keeping it improved the policy.
• Retraining on byte-identical data with only a different seed moved one LIBERO task's success rate by a standard deviation of 20.4 percentage points, and the ten-task average by 4.23. A calibration against 82 observational outcomes could not have measured what it claimed to, because the noise is the same size as the effect.

The mechanical parts of ORBIT were never the problem. Dead-joint detection, joint clipping, NaN and schema checks, temporal-alignment measurement and format conversion all report real, measurable properties of a dataset. What was wrong was the layer built on top of them that claimed to predict training outcomes.

What replaced it

The measurement programme that produced the retraction continues, and the honest part of the work — statistics for robot-policy evaluation, not predictions about datasets — moved to a separate package:

orbit-eval — noise floors, power analysis, correctly-paired comparisons, per-task release selection, and refusal logic that returns no number when a design cannot support one.

The full study is "Noise Floors and Non-Portable Value: What Robot Imitation Data Curation Can and Cannot Measure" (Lasne, 2026).

If you are still using this package

Pin orbit-robotics==0.7.2 if you depend on the loaders or the mechanical checks; they work and they are honest. Do not rely on any grade, prediction, "fit" score, sim2real transfer estimate, or episode-value output from any version — those are the retracted surfaces.

Why this release exists at all

A package that shipped a false claim for over a year should not simply vanish. Deleting the project would free the name for someone else and would read as hiding. This release exists so that anyone who installs orbit-robotics — or who finds it in an old requirements file — is told plainly what was wrong.

The research in this repository

This repo is also the measurement programme behind the retraction — the only replicate-retraining variance corpus in robot imitation learning, with the pre-registrations, the scored results, and the negative findings kept in.

• RETRAIN_LOTTERY.md — the flagship result and the 8-retrain measurement protocol
• research/ — pre-registered designs, scored results, freeze chains
• experiments/forecast/atlas.csv — the 16-regime variance atlas
• paper/DRAFT.md — Noise Floors and Non-Portable Value

Author

Rahil Lasne (@Rahillasne) — ORBIT Research.

License

MIT (unchanged).