Sigma_0: ACT-GAZER and DP-GAZER

24 student policies for six RoboTwin tasks, trained with teacher visual KL = 0 and student visual KL = 0.01, training seed 0.

Code: Cuixxx/Sigma_0 v1.0.0, commit e02e6325a87738491b2ca182bdd244da47ab162c.

Family	Teacher prior	Tasks	Checkpoint
ACT-GAZER	Gray-mask, all-camera	6	policy_best.ckpt
ACT-GAZER	Clean, all-camera	6	policy_best.ckpt
DP-GAZER-ATTN	Clean	6	600.ckpt
DP-GAZER-ATTN	Gray-mask	6	600.ckpt

Tasks: click_bell, place_empty_cup, place_container_plate, beat_block_hammer, lift_pot, stack_bowls_two.

DP uses native 240 x 320 head-camera observations resized to 480 x 640 by the encoder. Both teacher and student stages ran for 600 epochs. ACT uses three cameras, 6000 configured training epochs, and the best validation checkpoint. ACT's action-latent KL is a separate setting from visual KL.

Download

from huggingface_hub import snapshot_download
snapshot_download(repo_id="Lucascui97/Sigma_0", local_dir="sigma0_weights")

To download one model, use allow_patterns, for example:

snapshot_download(
    repo_id="Lucascui97/Sigma_0",
    local_dir="sigma0_weights",
    allow_patterns=[
        "DP-GAZER/gazer_distractor_strong-click_bell-aloha_agilex-joint-dp_gazer_attn-student-graymask-noteacherkl-res480-0/*",
        "manifest.json",
    ],
)

Loading and evaluation

Use the policy implementations and dependencies in the linked code release. These are custom policy checkpoints, not Transformers from_pretrained models.

• ACT: each experiment directory contains policy_best.ckpt, dataset_stats.pkl, and training_config.json, required by XPolicyLab/policy/ACT_GAZER/model.py. Configure ckpt_dir to that directory.
• DP: each experiment directory contains 600.ckpt, the original final training payload with config, workspace states, optimizer and EMA states when present. The runtime adapter in XPolicyLab/policy/DP/model.py chooses the largest numeric checkpoint by default. config.yaml is a readable export of the checkpoint's embedded config.
• The DP experiment launchers also check for latest.ckpt. Before using them, create a local alias in each downloaded DP experiment directory:

for directory in sigma0_weights/DP-GAZER/*; do
    if [ -f "$directory/600.ckpt" ] && [ ! -e "$directory/latest.ckpt" ]; then
        ln -s 600.ckpt "$directory/latest.ckpt"
    fi
done
export DP_GAZER_CHECKPOINT_ROOT="$PWD/sigma0_weights/DP-GAZER"

Training configs retain original provenance paths; update dataset and output paths for your environment. Teacher checkpoint files and training datasets are not included. Normal student inference uses RGB and robot state. Full training resumption may require the original data and teacher resources.

manifest.json lists each model, file size, and SHA-256 digest. Checkpoints have been checked against their stored training configuration; this upload does not represent a new simulation evaluation. Deployment on physical robots has not been validated by this release.