rldx2_pi05_robotwin_deltaeef_b128_60k

Vanilla baseline for the RLDX-2 project: RoboTwin 2.0 leaderboard co-train setting (50 tasks x 50 demo_clean episodes, Aloha-AgileX, official LeRobot export).

• Architecture: Pi0.5 (public lerobot 0.5.2 @ 09808183, chunk 50, three camera views, bf16 training)
• Optimizer batch 128, 60000 steps, final checkpoint (optimizer state not included)
• Inputs: head + left/right wrist images, 14-d joint state, the task instruction; outputs 14-d absolute joint targets
• Evaluation adapters: https://github.com/myungkyuKoo/RLDX-2 (ops/eval/xpl)

Configs reference the base backbone / tokenizer by hub id.

Intermediate checkpoints

Inference subset of every saved training step below the final (same layout as the repo root, optimizer state not included), one folder per step with its own SHA256SUMS:

• checkpoints/step-5000
• checkpoints/step-10000
• checkpoints/step-15000
• checkpoints/step-20000
• checkpoints/step-25000
• checkpoints/step-30000
• checkpoints/step-35000
• checkpoints/step-40000
• checkpoints/step-45000
• checkpoints/step-50000
• checkpoints/step-55000

One step: snapshot_download("Myungkyu/rldx2_pi05_robotwin_deltaeef_b128_60k", allow_patterns="checkpoints/step-5000/*"); the final alone: ignore_patterns="checkpoints/*".