KirkHaskell/act_delta_robot_bolt_pick_place

ACT: delta robot bolt pick-and-place (100 demos, 50k steps)

An ACT policy trained with LeRobot 0.5.0 on all 100 demonstrations of KirkHaskell/delta_robot_bolt_pick_place. It drives a home-built 3-DOF delta robot with an electromagnet gripper: it picks up a standing bolt and drops it into a cup.

Real-robot result: 6/10 successes (95% Wilson CI 31–83%). It was tied for the best ACT model in a 12-model ACT grid, and it was never stopped as unsafe. Its failures were 4 stalls or timeouts. It succeeded at both out-of-distribution trials (bolt outside the demonstrated area). Full study: https://github.com/KirkEHaskell/delta-robot-act-vs-diffusion

Inputs	observation.state (3 joint angles, deg) + 3 cameras (side, top_right, top_left, 240×320)
Output	4-D action: 3 absolute joint targets (deg) + magnet (on if > 0.5)
Settings	LeRobot ACT defaults: ResNet-18 (ImageNet), chunk size 100, all 100 actions executed, VAE (KL weight 10), LR 1e-5
Training	50k steps, batch 32, seed 1000, fp32, one RTX 3090 (~3.1 h)
Inference speed	0.1–0.37 s per 100-action chunk on a laptop CPU (Ryzen AI 9 HX 370), real time at 30 Hz

Use

hf download KirkHaskell/act_delta_robot_bolt_pick_place --local-dir models/act
cd robot && python run_policy.py dry --ckpt ../models/act

Normalization lives outside the policy in LeRobot 0.5.x. Load the bundled pre/post-processors with make_pre_post_processors(policy_cfg, pretrained_path=...), or the actions come out normalized.

Limitations: this model is specific to one robot, camera placement, scene and camera colour settings (two cameras have a fixed white balance). Expect to fine-tune on your own demonstrations.