mybotshop-act-v2

Trained ACT policy from the MyBotShop technical evaluation submission. Pick-and-place on a Franka Panda in PyBullet.

Source repo: https://github.com/NZ5253/ros2-il-pipeline

Results (20 / 50 closed-loop rollouts, RTX 4060)

Setup	Success
In-distribution (training cube spawn range)	48 / 50 = 96 %
Out-of-distribution (cube spawn shifted outside the training range)	1 / 20 = 5 %

Config

• 5.85 M params (LeRobot ACTPolicy, transformer encoder/decoder with VAE prior)
• Input split: STATE (joint pos + vel, 14-D) + ENV (EE pose + cube xyz, 10-D)
• Chunk size 50 (1.67 s planning horizon at 30 Hz)
• Trained 500 epochs, AdamW lr 1e-4, batch 32, KL weight 10
• Deployed with temporal ensembling (temporal_ensemble_coeff=0.01, n_action_steps=1)

Load it

import torch
ckpt = torch.load("best.pt", map_location="cuda:0", weights_only=False)
# state_dim 24, action_dim 7, chunk_size 50

Or through the inference node in the source repo:

ros2 service call /inference_node/load_policy il_pipeline_msgs/srv/LoadPolicy \
    "{checkpoint_path: 'best.pt', policy_type: 'act'}"

See il_pipeline/inference/policy_loader.py in the source repo for the full load + adapter path.