group2-lab3-interactiverobot/act_pick_doll

ACT Pick Doll — SO-ARM101

This repository contains a LeRobot Action Chunking Transformer (ACT) policy trained for the task “Grab the lion plushie.”

Model details

• Policy: ACT (ACTPolicy)
• Robot: SO-ARM101
• Observation: one 640×480 RGB image from observation.images.front and a six-dimensional joint state
• Action: six-dimensional joint command
• Action chunk: 100 steps
• Vision backbone: ResNet-18 with ImageNet initialization
• Parameters: 51,597,190
• Training: 20,000 optimizer steps, batch size 32, seed 100000
• Software: LeRobot 0.6.1, PyTorch 2.11.0, Python 3.12.13
• Dataset: 35 motion-integrity-curated episodes, 15,750 frames at 15 Hz

The saved policy was reloaded successfully with ACTPolicy.from_pretrained. The first and last logged training losses were 23.09 and 0.069 respectively.

Loading

import torch
from lerobot.policies.act.configuration_act import ACTConfig
from lerobot.policies.act.modeling_act import ACTPolicy
from lerobot.policies.factory import make_pre_post_processors

repo_id = "group2-lab3-interactiverobot/act_pick_doll"
config = ACTConfig.from_pretrained(repo_id)
config.device = "cuda" if torch.cuda.is_available() else "cpu"
config.use_amp = config.device == "cuda"
config.pretrained_backbone_weights = None  # All backbone weights are in the checkpoint.
policy = ACTPolicy.from_pretrained(repo_id, config=config, strict=True)
preprocessor, postprocessor = make_pre_post_processors(
    policy.config,
    pretrained_path=repo_id,
    preprocessor_overrides={"device_processor": {"device": config.device}},
)

Use the exact observation keys and shapes described in config.json. Robot deployment additionally requires matching calibration, camera placement, and a 15 Hz control loop.

Package verification

The packaged weights and processor state passed the supplied SHA-256 checksums. With LeRobot 0.6.1 and PyTorch 2.11.0, strict checkpoint loading and both processor pipelines were tested on CPU. A synthetic observation produced a finite action tensor with shape [1, 6]. This checks software loading and inference, not physical task performance.