You can deploy this model and start earning money today!

LogicDeposit/
Adamu_260331_RgbCube_act

Name: Adamu_260331_RgbCube_act
Availability: InStock

Views

No views yet

LogicMimic ACT Policy — AdamU 260331 RgbCube

A trained Action Chunking with Transformers (ACT) policy for the AdamU dual-arm humanoid robot. The model is trained using the LogicMimic framework (built on LeRobot) on multi-camera RGB observations.

Model Details

Architecture: ACT (Action Chunking with Transformers)
Vision Backbone: ResNet-18 (ImageNet1K V1 pre-trained)
Framework: LeRobot / PyTorch
Training Device: CUDA
Training Checkpoint: 61,800 steps

Input

observation.state: [16] (selected joint positions)
observation.images.head: [3, 360, 640] (RGB)
observation.images.wrist_left: [3, 480, 640] (RGB)
observation.images.wrist_right: [3, 480, 640] (RGB)

Output

action: [16] (joint commands)

Hyperparameters

Parameter	Value
Chunk Size	50
Action Steps	50
Optimizer	AdamW
Learning Rate	1e-05
Backbone LR	1e-05
Weight Decay	0.0001
Batch Size	32
KL Weight	10.0
Dropout	0.1
Encoder Layers	4
Decoder Layers	1
Model Dim	512
Attention Heads	8
Latent Dim (VAE)	32
Normalization	MEAN_STD

Training Data

Dataset: LogicDeposit/Adamu_260331_RgbCube
Valid Episodes: 0–99 (100 episodes)
Total Frames: 49,415
FPS: 30 Hz
WandB Project: pnd-learning (run ID: 2dfgw37r)

Deployment

To deploy this policy, use the LogicMimic inference pipeline.

Synchronous Inference (Single Machine)

bash

python scripts/launch_sync.py \
    --model_path <MODEL_PATH> \
    --dataset_path <DATASET_PATH> \
    --camera_source zmq \
    --zmq_camera_host <NX_IP> \
    --device mps \
    --robot_ip <NUC_IP> \
    --fps 10 \
    --task 'pick the bottle and place into the box' \
    --temporal_ensemble_coeff 0.01

Asynchronous Inference (Recommended)