Myungkyu/rldx_1_rmbench_taco_b64_60k

rldx1rmbenchtacob64_60k

Low-level policy for RMBench (9 simulated tabletop tasks), fine-tuned from RLWRLD/RLDX-1-PT on `Myungkyu/RMBench-taco-gemini` — RMBench demonstrations with dense subtask labels from the task-specific context (offline annotation).

• —Architecture: RLDX-1-PT (video length 4, three camera views + one keyframe slot)
• —Optimizer batch 64, 60000 steps, final checkpoint (the run was launched as batch 128 = 64 x gradient accumulation 2; the DeepSpeed ZeRO-2 build applied the accumulation without the intended scaling, so the effective optimizer batch was 64 — verified from the gradient norms on 2026-09-11; renamed from rldx_1_rmbench_taco_b128_60k on 2026-09-14)
• —Inputs: head + left/right wrist images, proprioception, the current subtask text; the keyframe slot takes a retrieved past frame when the label calls for one

Configs reference the base backbone / tokenizer by hub id or by the training site's local path — point them at your local copies before loading.