Skip to content

[Performance] Reduce repeated static context in Pilot planning rounds #209

Description

@enkerewpo

Problem

Pilot rebuilds a large prompt each planning round, including stable RTDL protocol text, Soma/URDF embodiment data, capability descriptions, conversation history, and live state. In the preliminary first-12 benchmark, Pilot calls had a median observed input of 25,829 tokens and a median prompt text size of about 45.3 KiB; even the single successful Robonix cell required 37 model calls.

Official and OM1 use much smaller planner-facing state/action representations, so their model budget is more directly spent on task decisions.

Acceptance criteria

  • Measure prompt bytes/tokens by section for each planning round.
  • Cache or reference stable protocol, capability, and embodiment sections instead of reconstructing/resending unchanged content.
  • Send delta live-state/feedback where possible while preserving plan/call provenance.
  • Add a representative multi-step regression that preserves task behavior while reducing median Pilot input tokens and end-to-end calls.
  • Publish before/after metrics using the same frozen protocol and model.

The cited first-12 results are preliminary diagnostics. Robonix is under active optimization, and these results do not represent final system performance.

Metadata

Metadata

Assignees

Labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions