Python / ManiSkill / LeRobot ACT
LangMani
Typed language-to-task routing and six task-specific controllers in a frozen simulation protocol.
Private source; access required · Version pinned to the evidence below
Work and recorded results
- Built typed language-to-task routing and six task-specific ACT controllers in simulation, retaining rejected requests and controller/runtime failure attribution.
Scope of the evidence
- The v1 protocol planned 72 pairs: 61 language requests executed and 11 were rejected. Results are 46/72 versus 55/72, not 72 physically executed state pairs.
- The v1 tag is distinct from the older default README. Later SmolVLA Phase 2C-C results (1/30 train resets; 0/30 validation; unseen not run) do not constitute an improvement to v1. Private source access is required.
Source and evidence
Evidence source revision: 58434cb17a7234b6d4b2c4fb15aecf8df0621487
Private source; access required at this revision ↗
Updated 2026-09-14. CV wording and this page use the same project manifest.