QZWang/patchwam-robotwin-c2r-step140000

PatchWAM — RoboTwin 2.0 C2R (step 140000)

This repository contains the checkpoint used for the reported PatchWAM RoboTwin 2.0 C2R result.

Result

Condition	Success rate	Protocol
Clean	91.56%	50 tasks × 100 episodes
Randomized	66.72%	50 tasks × 100 episodes
Average	79.14%	10,000 rollouts total

The full task-level result is included as results/c2rdr4_140k_100ep_full.json. The average is the arithmetic mean of the Clean and Randomized task means.

Checkpoint

• File: step_140000.pt
• Size: 7,751,411,897 bytes
• SHA256: 22f556d02903a06f23f21a8a28ef91ff92611de0bad4efbf8827c01bd3701fd9
• Contents: model weights, step number, dtype, and proprio encoder weights; no optimizer state or training paths
• Architecture: FLUX.2 Klein 4B Action-as-Patch
• Initialization: FLUX.2 Klein base 4B (no prior Action-as-Patch pretraining)
• Action/state dimension: 14
• Action horizon / replan: 16 / 16
• Evaluation denoising steps: 10

huggingface-cli download QZWang/patchwam-robotwin-c2r-step140000 step_140000.pt \
  --local-dir ./checkpoints
sha256sum ./checkpoints/step_140000.pt

Training recipe

The portable recipe is in training_recipe.yaml. The run used 2 nodes × 8 GPUs, per-GPU batch 16, gradient accumulation 1, and global batch 256. The evaluated checkpoint is step 140,000.

Training data statement

Training used only the RoboTwin 2.0 training data selected by the C2R recipe recorded in training_recipe.yaml: periodic_prefix(period=550, keep_first=50) followed by the recorded non-idle filter. The resolved run configuration contains a single RoboTwin 2.0 dataset source. No evaluation rollouts or private additional datasets were used for training.

Users are responsible for complying with the upstream RoboTwin, FLUX.2, autoencoder, and text-encoder licenses.

Evaluation

• Benchmark: RoboTwin 2.0 C2R
• Tasks: 50
• Conditions: Clean and Randomized
• Episodes: 100 per task per condition
• Total rollouts: 10,000
• Action horizon: 16
• Replan interval: 16

See metrics.json and the full task-level JSON under results/.