Python / LeRobot / asynchronous inference

ActionStream

Episode-owned asynchronous policy execution with cancellable transport, revision-tagged actions and measured X-VLA/LLM development tooling.

Private source; access required · PR 3 merged into master; experimental language and vision gates failed

Work and recorded results

Built a LeRobot-compatible asynchronous policy runtime with episode-owned queues, cancellable process transport and revision-tagged action delivery; repaired 11 reproduced lifecycle failures and added native X-VLA/LIBERO development tooling.
Integrated native X-VLA action processing and original-request evidence binding; retained a four-example simulation baseline and explicit failed language/vision quality gates.

PR 3 is merged; GitHub CPU core/quality and upstream LeRobot contract checks passed at this revision. This refresh read the evidence and CI; it did not rerun GPU models or physical robots.
One smoke and three development resets on a single LIBERO task succeeded, with 514 control steps and 20 inference calls on an RTX 4090. These four examples are not a generalization rate; EGL used software rendering.
The new 48-case language comparison still falsely admitted one of 24 negative requests. All 20 visual outputs violated the JSON-only contract; accepted coverage was 0% on 16 labelled observations. Both quality gates failed; no integrated language/vision robot-control episode was run.
Stage-level CPU test counts overlap and must not be summed. Historical H1-R2 supply/cost tradeoffs and H2 NO-GO remain unchanged; the upstream LeRobot proposal is not merged. Source remains private.

Reviewed source revision: 5e847811ded4d62a8ed403c6c000f91c012f4807

Updated 2026-09-14. CV wording and this page use the same project manifest.