bench2dex-wuji-policies

Bench2Dex Wuji policy experiments

Checkpoints from independent single-GPU runs on the Condiment Box Loading and Ball Box Loading tasks. The robot action space is 52 absolute joint targets. See artifacts/results-table.md for measured evaluation evidence.

Methods: Bench2Dex ACT/DP and DECO vision. Current comparison runs use physical batch 32, accumulation 1 and 80000 optimizer updates per task. Historical epoch-recipe and diagnostic checkpoints remain separately named and are not interchangeable with this comparison. Additional methods include VITA, Dexora-policy (native internal 384 vision, without the full quality-filtered posttraining), and SPD-style (explicit paper-architecture reimplementation without original SPD robot pretraining). Different native histories, visual pretraining and sequence-label exposure remain documented in each run config; equal B32/80k does not mean equal data exposure or compute. Runtime-recovery checkpoints include optimizer state and are not all final selected policies. Raw demonstrations are provided by Bench2Dex, not redistributed here.