Diffusion Policy · dual OpenYAM, "close the box"

A Diffusion Policy policy for a bimanual OpenYAM robot, trained on kiteml/dual-openyam-close-box to close the box. Diffusion Policy is a visuomotor policy that generates action sequences by iterative denoising. Trained from scratch.

Trained with Kite on LeRobot v0.6.0.

Training


Dataset	`kiteml/dual-openyam-close-box`: 31 episodes, 33,247 frames at 30 fps
Initialization	none (from scratch)
Trained	all parameters (ResNet-18 backbone initialized from ImageNet)
Steps	50,000
Batch size	8
Optimizer	Adam, lr 1e-4, cosine schedule with warmup
Final training loss	0.004
Hardware	1× NVIDIA A100 40GB

Inputs and outputs

Feature	Shape	Contents
`observation.images.top`	3×480×640	top camera, RGB
`observation.images.left_wrist`	3×480×640	left wrist camera, RGB
`observation.images.right_wrist`	3×480×640	right wrist camera, RGB
`observation.state`	14	left arm joints 1–6, right arm joints 1–6, left gripper, right gripper
`action`	14	same layout as `observation.state`

Each inference call predicts 64 actions (2.1 s at 30 fps) and executes 32 of them, conditioned on the last 2 observations.

Usage

Requires LeRobot v0.6.0 or later.

import torch
from lerobot.policies.factory import get_policy_class, make_pre_post_processors

repo_id = "kiteml/dual-openyam-close-box-diffusion"
policy = get_policy_class("diffusion").from_pretrained(repo_id)
preprocessor, postprocessor = make_pre_post_processors(policy.config, pretrained_path=repo_id)

# observation: the input features above, formatted like a LeRobotDataset frame
# (images as float tensors in [0, 1], channels first)
observation["task"] = "close the box"
with torch.inference_mode():
    action = postprocessor(policy.select_action(preprocessor(observation)))

To run it with LeRobot's CLI tools, pass --policy.path=kiteml/dual-openyam-close-box-diffusion to lerobot-record or lerobot-eval.