act-ocbench-blue-cube

ACT policy (LeRobot implementation) trained on Autobrik/ocbench-blue-cube-pnp: blue-cube pick-and-place in OCBench (MuJoCo, UR5e + Robotiq 2F-85) from three 224x224 RGB cameras and 7-D robot proprioception. Actions are absolute joint and gripper-driver targets at 50 Hz. Code, data contract and evaluation protocol: https://github.com/BrikHMP18/act-weight-editing

The repository root holds the checkpoint at step 60000, selected by the lowest logged validation loss on held-out episodes. All saved checkpoints are under checkpoints/<step>/. Validation loss is a proxy; simulation success rates are reported in the code repository.

Checkpoints

Step	Validation loss
20000	0.3972
40000	0.3919
60000	0.3847	selected
80000	0.3978
100000	0.3874

Training settings

Setting	Value
chunk_size	100
n_action_steps	1
temporal_ensemble_coeff	0.01
dim_model	512
n_encoder_layers	4
n_decoder_layers	1
kl_weight	10.0
optimizer_lr	1e-05
vision_backbone	resnet18
batch_size	8
steps	100000
eval_split	0.2
seed	1000

Usage

from lerobot.policies.act.modeling_act import ACTPolicy
from lerobot.policies.factory import make_pre_post_processors

policy = ACTPolicy.from_pretrained("Autobrik/act-ocbench-blue-cube")
pre, post = make_pre_post_processors(policy.config, pretrained_path="Autobrik/act-ocbench-blue-cube")

Evaluate in simulation with the code repository's act-rollout --checkpoint <local dir>.