RoboActualizer checkpoints

Checkpoints of Roboactualizer, a joint image/action flow-matching policy on a frozen V-JEPA 2.1 ViT-L/16 encoder. Code, environments and full instructions: the Roboactualizer repository.

Folder	Model	Benchmark	Inference	Result	Eval config
libero_dits/	DiT-S, step 173580	LIBERO (4 suites x 10 tasks x 50)	4 steps, no CFG, bf16	98.00	eval_libero_dits
libero_dits/	DiT-S, step 173580	LIBERO-Plus (10,030 tasks x 1)	4 steps, action text-CFG 1.5, bf16	63.1	eval_libero_plus_dits
libero_ditb/	DiT-B, step 130185	LIBERO-Plus (10,030 tasks x 1)	4 steps, action text-CFG 1.5, bf16	65.0	eval_libero_plus_ditb
robotwin_dits_fp32/	DiT-S, fp32 weights, step 176540	RoboTwin 2.0 (50 tasks x clean/randomized x 25)	10 steps, no CFG, fp32	58.84	eval_robotwin_dits

Each folder holds the checkpoint, the dataset_stats.json used for action/state normalization (keep it next to the checkpoint; evaluation looks for it there) and the training config.yaml. SHA256SUMS lists file hashes.

All checkpoints use the 300M V-JEPA 2.1 encoder: https://dl.fbaipublicfiles.com/vjepa2/vjepa2_1_vitl_dist_vitG_384.pt, saved as checkpoints/vjepa2/vjepa2_1_vit_large_384.pt in the code repository.

Usage

huggingface-cli download db12312607/RoboActualizer --exclude "data/*" --local-dir checkpoints/roboactualizer

# from the code repository root
python experiments/libero/run_libero_manager.py --config-name eval_libero_dits \
    ckpt=checkpoints/roboactualizer/libero_dits/step_173580.pt
bash experiments/libero/run_libero_plus.sh --config-name=eval_libero_plus_ditb \
    ckpt=checkpoints/roboactualizer/libero_ditb/step_130185.pt
python experiments/robotwin/run_robotwin_manager.py --config-name eval_robotwin_dits \
    ckpt=checkpoints/roboactualizer/robotwin_dits_fp32/step_176540.pt