YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
RMBench++ battery_test_sort adaptive 16v3 P2/P3
Final step 2,000 for both phases: P2 joint training, then P3 memory-only training initialized from P2. Both use LR 1e-4 and completed successfully at 2026-10-16T06:17:20Z.
RoboDojo dual X5 absolute 14D actions; native measured proprioception; cameras cam_head, cam_left_wrist, cam_right_wrist; 160x192 mosaic. Adaptive delta_pointwise_threshold router, GIST4/recent FULL4, eight slots/eight region tokens, four memory denoising steps, quantile gate 0.90 and prefix limit 60.
Training code 7a60935509a7a56aef7b62dc7a39c040a2a60d29; published equivalent implementation: sgres-paw/Little-WAM research-marina/v16.3_memory_tokens_speed_opt at 7a60935509a7a56aef7b62dc7a39c040a2a60d29.
Dataset: original battery-test-sort RoboDojo export, 50 demonstrations, 38,172 frames, 25Hz. Export SHA256: 687150e278eed028bf3b110a47534a1b85aa10cfbcccb326532ff18ad1666414. Base checkpoint: step_016428.pt, SHA256 9a9919146af4042ecb9c0756fb05efb6292cea94b084f729f4d451773cf295c7. Its 48-chunk supervision window was explicitly migrated to full episodes with new 16v3 slots; the original action, observation and normalization contracts are preserved.
Layout matches m1b/kairos_rmbenchpp_rearrange_blocks_interrupt_16v3_p2p3: each phase contains step_002000.pt, train_args.json, normalization_stats.json and dataset_summary.json. latest.json is omitted. MANIFEST.json binds artifacts by byte size and SHA256. No policy evaluation score is claimed.