Mohit Shridhar*, Yat Long (Richie) Lo*, Stephen James
CoRL 2024

Genima fine-tunes Stable Diffusion to draw joint-actions on RGB observations.

This repo is for reproducing the RLBench results from the paper. For the latest updates, see: genima-robot.github.io.

• Getting Started: Installation, Quickstart, Checkpoints and Pre-Generated Datasets, Model Card
• Training & Evaluation: Full Pipeline, Colosseum Tests, Other Diffusion Base Models
• Miscellaneous: Notebooks, Disclaimers, FAQ, Licenses
• Acknowledgements: Acknowledgements, Citations

Genima is built with Python 3.10.12. We use poetry to manage dependencies.

cd <install_dir>
conda create -p genima_env python==3.10.12        # create conda env
conda activate genima_env                         # activate env

pip install poetry
poetry self add poetry-exec-plugin                # install plugin for executables
poetry self update

cd <install_dir>
git clone https://github.com/MohitShrdhar/genima.git
cd genima
poetry exec rlbench                               # install pyrep and rlbench
poetry install                                    # install dependencies

Manually install RoboBase (which contains the ACT implementation) for debugging purposes:

cd <install_dir>
git clone https://github.com/robobase-org/robobase.git
cd robobase
pip install -e .

This is a quick tutorial on evaluating a pre-trained Genima agent.

Download the pre-trained checkpoint trained on 25 RLBench tasks with 50 demos per task:

cd genima
poetry exec quick_start

Generate a small val set of 10 episodes for open_box inside /tmp/val_data:

mkdir /tmp/val_data
cd genima/rlbench/tools
python dataset_generator.py \
     --save_path=/tmp/val_data \
     --tasks=open_box \
     --image_size=256,256 \
     --renderer=opengl \
     --episodes_per_task=10 \
     --variations=1 \
     --processes=1 \
     --arm_max_velocity 2.0 \
     --arm_max_acceleration 8.0

Evaluate the pre-trained Genima agent:

cd genima/controller
python eval_genima.py \
     task=open_box \
     dataset_root=/tmp/val_data \
     diffusion_ckpt=../ckpts/25_tasks/diffusion_sdturbo_R256x4_tiled \
     controller_ckpt=../ckpts/25_tasks/controller_act \
     num_eval_episodes=10 \
     save_gen_images=False \
     num_diffusion_steps=5 \
     execution_horizon=20 \
     save_video=False \
     wandb.use=False \
     eval_type=latest \
     headless=False

If you are on a headless machine, turn off RLBench visualization with headless=True. You can save the generated target images to /tmp/ by setting save_gen_images=True.

You can evaluate the same Genima agent on other tasks by generating a val set for that task.

We provide pre-trained checkpoints for RLBench agents:

• 25 Task Genima

• 3 Task Genima - from ablations

See quickstart on how to evaluate these checkpoints.

TBD

This guide covers how to train Genima from scratch.

Use the dataset_generator.py tool to generate datasets:

cd rlbench/tools

# generate train set
python dataset_generator.py \
     --save_path=/tmp/train_data \
     --tasks=take_lid_off_saucepan \
     --image_size=256,256 \
     --renderer=opengl \
     --episodes_per_task=25 \
     --variations=1 \
     --processes=1 \
     --arm_max_velocity 2.0 \
     --arm_max_acceleration 8.0


# generate val set
python dataset_generator.py \
     --save_path=/tmp/val_data \
     --tasks=take_lid_off_saucepan \
     --image_size=256,256 \
     --renderer=opengl \
     --episodes_per_task=10 \
     --variations=1 \
     --processes=1 \
     --arm_max_velocity 2.0 \
     --arm_max_acceleration 8.0

Note: If you have old RLBench datasets, they won't work with Genima. You need RLBench master up until this commit to save joint poses.

To render actions, provide paths to your RLBench dataset and random textures:

# downloads textures for random backgrounds
poetry exec download_textures

# use pyrender to place spheres that at joint-actions that t+20 timesteps ahead
cd render
python render_data.py \
     episodes=25 \
     dataset_root=/tmp/train_data \
     textures_path=./mil_textures/object_textures \
     action_horizon=20 \
     num_processes=5

By default, two dataset folders are generated: rlbench_data_rgb_rendered with observations and joint targets to train the diffusion agent, and rlbench_data_rnd_bg with random backgrounds and joint targets to train the controller. See the sample notebook for visual illustrations of the rendered data.

# setup your accelerate
accelerate config

# finetune SD-turbo with controlnet
cd diffusion
python train_controlnet_genima.py
     --pretrained_model_name_or_path='stabilityai/sd-turbo' \
     --output_dir=/tmp/diffusion_agent \
     --resolution=512 \
     --learning_rate=1e-5 \
     --data_path='/tmp/train_data_rgb_rendered/' \
     --validation_images_path '/tmp/train_data_rgb_rendered' \
     --train_batch_size=2 \
     --checkpoints_total_limit=2 \
     --num_train_epochs=100 \
     --report_to wandb \
     --report_name 'sdturbo_1task_R256x4_tiled' \
     --image_type 'tiled_rgb_rendered' \
     --conditioning_image_type 'tiled_rgb' \
     --tasks 'take_lid_off_saucepan' \
     --validation_steps 500 \
     --mixed_precision='fp16' \
     --variant='fp16' \
     --allow_tf32 \
     --enable_xformers_memory_efficient_attention \
     --tiled \
     --num_validation_images 1 \
     --augmentations=crop,colorjitter \
     --num_demos 25 \
     --checkpointing_steps 1000 \
     --resume_from_checkpoint 'latest'

Monitor the training on wandb to check the quality of the generated targets. If the spheres are blurry, at the wrong location, or have the wrong color, then the model is not trained enough. You need train between 100-200 epochs for good results. For multi-task training just provide a comma-separated list --tasks 'take_lid_off_saucepan,open_box'.

# train ACT to map target images to a sequence of joint-actions
cd controller
python train_act.py \
     env=rlbench \
     env.dataset_root=/tmp/train_data_rnd_bg/ \
     work_dir=/tmp/controller \
     demos=25 \
     env.train_tasks=[take_lid_off_saucepan] \
     num_train_epochs=1000 \
     action_sequence=20 \
     batch_size=8 \
     method.lr=1e-5 \
     wandb.use=true

The ACT controller for Genima can be trained independently of the diffusion agent. If you have sufficient compute, you train both the diffusion agent and controller simultaneously.

The hyperparameters of the controller are set in controller/cfgs/method/genima_act.yaml. For multi-task training just provide a comma-separated list env.train_tasks=[take_lid_off_saucepan,open_box]'.

To train the ACT baseline, set env.dataset_root=/tmp/train_data to use raw RGB observations instead of target spheres with random backgrounds. See the RoboBase repository for other baselines.

# Use the diffusion agent and controller sequentially to evaluate
python eval_genima.py \
     task=take_lid_off_saucepan \
     dataset_root=/tmp/val_data \
     diffusion_ckpt=/tmp/diffusion_agent/sdturbo_1task_R256x4_tiled \
     controller_ckpt=/tmp/controller \
     num_eval_episodes=10 \
     save_gen_images=False \
     num_diffusion_steps=5 \
     execution_horizon=20 \
     save_video=True \
     wandb.use=True \
     eval_type=last_three \
     headless=True

To run the evaluation offline, set headless=False. By setting eval_type=last_three, the script will sequentially evaluate the last three checkpoints and report average scores. Alternatively, you can set eval_type=latest or eval_type=980 for specific checkpoints.

You can visualize the generated targets by setting save_gen_images=True. This will save the diffusion outputs to /tmp. However, note that saving images to disk is slow.

For the fastest inference speed, set torch_compile=True and enable_xformers_memory_efficient_attention=False. See other optimizations here.

All RLBench experiments in the paper use num_diffusion_steps=10, execution_horizon=20, num_eval_episodes=50, and eval_type=last_three.

You can evaluate the same checkpoints from quickstart on 6 pertubation categories from Colosseum.

python eval_genima.py \
     task=open_drawer \
     dataset_root=/tmp/val_data \
     diffusion_ckpt=/tmp/diffusion_agent/sdturbo_1task_R256x4_tiled \
     controller_ckpt=/tmp/controller \
     save_gen_images=False \
     num_eval_episodes=10 \
     save_video=True \
     wandb.use=True \
     eval_type=last_three \
     headless=True \
     colosseum_use=True \
     colosseum_task_config=cfgs/colosseum/random_object_color.yaml

Select from 6 config files for colosseum_task_config:

• Randomized object and part colors: cfgs/colosseum/random_object_color.yaml for open_drawer.
• Distractor objects: controller/cfgs/colosseum/distractor_objects.yaml for open_drawer.
• Lighting variations: controller/cfgs/colosseum/lighting_variations.yaml for open_drawer.
• Randomized background textures: controller/cfgs/colosseum/random_background_textures.yaml for move_hanger.
• Randomized table textures: controller/cfgs/colosseum/random_table_textures.yaml for basketball_in_hoop.
• Camera pose perturbations: controller/cfgs/colosseum/random_camera_poses.yaml for move_hanger.

• Colab Tutorial (soon): How to generate the dataset and train the diffusion agent.
• Dataset Visualizer: Looks into the rendered joint-target datasets.

• Parallelization: A lot of things (data generation, evaluation) are slow because everything is done serially. Parallelizing these processes will save you a lot of time.
• Variance in Success Rates: You may notice small variations in the success rate due to the stochastic nature of the simulator.
• Other Limitations: See the "Limitations and Potential Solutions" section in the paper appendix.

How long should I train for?

100-200 epochs for the diffusion agent. 1000 epochs for the controller.

How many training demos do I need?

It depends on the number, complexity, and diversity of tasks. Start with 50 demos in simulation and iteratively reduce the number demos until you achieve >80% of the peak performance.

Is multi-gpu training supported?

Yes for the diffusion agent, since it's based off HF diffusers. But no for the controller, since RoboBase only supports single-GPU training. You can use other ACT implementations to train the controller.

Will the real-robot code be released?

The Genima part of the real-robot code is identical to this repo. You just need format your dataset into the RLBench dataset format.

Will the real-world checkpoints and data be released?

No. Without our particular camera setup, these real-world datasets and checkpoints are not useable.

• Diffusion Agent Training: A100 with 80GB VRAM
• Controller Training: L4 with 24GB VRAM, 120GB RAM
• Evaluation: L4 or RTX 3090 with 24GB VRAM

Only the diffusion agent training requires GPUs with larger VRAMs. Both inference and controller training can be done on commodity GPUs.

Update 28-Aug-2024:

• Initial code release.

• Support LeRobot's ACT for the controller (needs FiLM conditioning)
• Host dataset on HF
• Host pre-trained checkpoints on HF

• Genima License (Apache 2.0) - This repository
• HuggingFace Diffusers License (Apache 2.0) - Fine-tuning stable diffusion code
• ACT License (MIT) - Original ACT code
• MT-ACT License (MIT) - Original multi-task ACT code
• RLBench Licence - Simulator
• PyRep License (MIT) - Simulator wrapper
• CLIP License (MIT) - Language encoder

Special thanks to Huggingface for Diffusers, Zhao et al. for the ACT repo, and Bharadhwaj et al. for the MT-ACT repo.

Genima

@inproceedings{shridhar2024generative,
  title     = {Generative Image as Action Models},
  author    = {Shridhar, Mohit and Lo, Yat Long and James, Stephen},
  booktitle = {Proceedings of the 8th Conference on Robot Learning (CoRL)},
  year      = {2024},
}

Diffusers

@misc{von-platen-etal-2022-diffusers,
  author = {Patrick von Platen and Suraj Patil and Anton Lozhkov and Pedro Cuenca and Nathan Lambert and Kashif Rasul and Mishig Davaadorj and Dhruv Nair and Sayak Paul and William Berman and Yiyi Xu and Steven Liu and Thomas Wolf},
  title = {Diffusers: State-of-the-art diffusion models},
  year = {2022},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huggingface/diffusers}}
}

ACT

@inproceedings{zhao2023learning,
  title={Learning fine-grained bimanual manipulation with low-cost hardware},
  author={Zhao, Tony Z and Kumar, Vikash and Levine, Sergey and Finn, Chelsea},
  booktitle = {Robotics: Science and Systems (RSS)},
  year={2023}
}

MT-ACT

@misc{bharadhwaj2023roboagent,
  title={RoboAgent: Generalization and Efficiency in Robot Manipulation via Semantic Augmentations and Action Chunking},
  author={Homanga Bharadhwaj and Jay Vakil and Mohit Sharma and Abhinav Gupta and Shubham Tulsiani and Vikash Kumar},
  year={2023},
  eprint={2309.01918},
  archivePrefix={arXiv},
  primaryClass={cs.RO}
}

Colosseum

@inproceedings{pumacay2024colosseum,
  title     = {THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulation},
  author    = {Pumacay, Wilbert and Singh, Ishika and Duan, Jiafei and Krishna, Ranjay and Thomason, Jesse and Fox, Dieter},
  booktitle = {Robotics: Science and Systems (RSS)},
  year      = {2024},
}