Learning & localization / SELECTED PROJECT

GOAI Bimanual Policy Training and Evaluation

Multicamera vision-language-action policy adaptation, training deployment, local closed-loop simulation and failure analysis.

My role: Individual competition project · Open-model adaptation, experiment organization and policy deployment

SYSTEM / FLOW

System schematic · not a device photo or execution record

00 / Context & objective

Multicamera vision-language-action policy adaptation, training deployment, local closed-loop simulation and failure analysis.

01My contributions

Adapted head and wrist camera inputs, language instructions, and 14-dimensional robot state and action interfaces.
Configured action-head fine-tuning and later multitask training, managing checkpoints, normalization settings and frozen versions.
Connected a WebSocket policy service to multiple simulation environments and organized recovery checks and evaluation guardrails.
Organized two full evaluations, candidate-checkpoint screening and task-level failure analysis with traceable results and video indices.

The stable baseline achieved 33/600 and the 20k fine-tune 28/600 across 24 tasks each: 1,200 episodes in total.
Candidate C checkpoints at 60k, 80k and 100k received 45 screening episodes in total, each scoring 0/15; no full 600-episode reevaluation was completed.
Diagnostic material covers 29 representative episodes and an index of 87 three-camera videos.

Read results alongside their measurement targets and conditions. Historical records, host checks, and device validation are distinguished.

StarVLA, QwenOFT, XPolicyLab and RoboDojo provide the upstream foundations. My work is adaptation, configuration, integration, runtime fixes and analysis.
These are local simulation records without a confirmed official score. The 20k aggregate observation is below baseline and does not establish improvement.
The 45-episode screening is not directly rankable against full 600-episode evaluations; no performance advantage is established for the 100k candidate.
Archiving did not rerun GPU training or full simulation. Raw 20k episodes, complete ACT source and implementations for some historical tests remain incomplete.