Add Aloha ACT notebook

notebooks/aloha_act/README.md

@@ -0,0 +1,102 @@
+# Imitation Learning with ACT (Action Chunking Transformer)
+
+This notebook demonstrates how to use the ACT (Action Chunking Transformer) model for imitation learning tasks for Aloha Robot, with OpenVINO optimization for efficient inference.
+
+**Note:** This notebook currently supports **Ubuntu 22.04** only.
+
+## Overview
+
+Imitation learning is a machine learning approach where a model learns to mimic expert behavior by observing and replicating demonstrations. ACT is an action chunking policy that uses Transformers for sequence modeling and is trained as a conditional VAE (CVAE) to capture variability in human data. It significantly outperforms previous imitation learning algorithms on various simulated and real-world fine manipulation tasks.
+
+For more details, see the original paper: [Action Chunking Transformer](https://arxiv.org/pdf/2304.13705)
+
+## Contents
+
+This notebook covers:
+
+1. **Prerequisites** - Downloading pre-trained weights
+2. **Dependency Installation** - Setting up required packages and repositories
+3. **Model Conversion** - Converting PyTorch checkpoint to OpenVINO IR format
+4. **Device Selection** - Choosing target hardware for inference
+5. **Policy Evaluation** - Running the trained policy on simulation tasks
+
+## Requirements
+
+- OpenVINO 2025.4.1
+- PyTorch 2.7.1
+- MuJoCo 3.2.6
+- Additional dependencies (automatically installed by the notebook)
+
+## Pre-trained Weights
+
+Pre-trained weights for the `sim_insertion_scripted` task are required. The notebook will automatically download and extract them from:
+- [Download Link](https://eci.intel.com/embodied-sdk-docs/_downloads/sim_insertion_scripted.zip)
+
+## Installation
+
+The notebook handles all installation automatically, including:
+
+- Cloning the edge-ai-suites repository (sparse checkout)
+- Setting up the ACT pipeline from the robotics-ai-suite
+- Installing DETR (Detection Transformer) dependencies
+- Applying OpenVINO-specific patches
+- Installing all required Python packages
+
+## Usage
+
+Simply run the notebook cells in order:
+
+1. **Download Pre-trained Weights** - Downloads and extracts model checkpoint
+2. **Install Dependencies** - Sets up environment and installs packages
+3. **Convert Model** - Converts PyTorch checkpoint to OpenVINO IR format
+4. **Select Device** - Choose CPU, GPU, or other available hardware
+5. **Evaluate Policy** - Run inference and evaluate the policy performance
+
+## Model Conversion
+
+The conversion uses the following parameters:
+```bash
+python3 ov_convert.py --ckpt_path <path> --height 480 --weight 640 --camera_num 4 --chunk_size 100
+```
+
+## Evaluation
+
+The evaluation script runs with these default parameters:
+```bash
+python3 imitate_episodes.py \
+  --task_name sim_insertion_scripted \
+  --ckpt_dir <path> \
+  --policy_class ACT \
+  --kl_weight 10 \
+  --chunk_size 100 \
+  --hidden_dim 512 \
+  --batch_size 8 \
+  --dim_feedforward 3200 \
+  --num_epochs 2000 \
+  --lr 1e-5 \
+  --seed 0 \
+  --device CPU/GPU \
+  --eval
+```
+
+## Repository Structure
+
+```
+aloha_act/
+├── aloha-act.ipynb           # Main notebook
+├── README.md                  # This file
+├── sim_insertion_scripted/    # Downloaded pre-trained weights
+│   └── four_camera/
+│       └── policy_best.ckpt
+└── edge-ai-suites/            # Cloned ACT pipeline
+    └── robotics-ai-suite/
+        └── pipelines/
+            └── act-sample/
+```
+
+## References
+
+- [ACT Paper](https://arxiv.org/pdf/2304.13705) - Action Chunking Transformer
+- [OpenVINO Toolkit](https://docs.openvino.ai/)
+- [Edge AI Suites](https://github.com/open-edge-platform/edge-ai-suites)
+- [Original ACT Implementation](https://github.com/tonyzhaozh/act)