thu-ml/RoboticsDiffusionTransformer

RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation

📝Paper | 🌍Project Page | 🤗Model | 🛢️Data | 🏞️Poster

RDT-1B is a 1B-parameter (largest to date) imitation learning Diffusion Transformer pre-trained on 1M+ (largest to date) multi-robot episodes. Given language instruction and RGB images of up to three views, RDT can predict the next $64$ robot actions. RDT is inherently compatible with almost all kinds of modern mobile manipulators, from single-arm to dual-arm, joint to EEF, position to velocity, and even with wheeled locomotion.

We have fine-tuned RDT on 6K+ (one of the largest) self-collected bimanual episodes and deployed it on the ALOHA dual-arm robot. It has achieved state-of-the-art performance in terms of dexterity, zero-shot generalizability, and few-shot learning. You can find Demo videos on our project page.

This repo is an official PyTorch implementation of RDT, containing:

• 🛠️Model implementation of RDT
• 🤗1M-step checkpoint of RDT-1B pre-trained on multi-robot data
• 🤗500K-step checkpoint of RDT-170M (RDT(small) in ablation)
• 📈Training and sampling scripts (with DeepSpeed)
• 🤖An example of real-robot deployment
• 🕹️Simulation benchmark from Maniskill environment

The following guides include the installation, fine-tuning, and deployment. Please refer to pre-training for a detailed list of pre-training datasets and a pre-training guide.

📰 News

• [2025/04/04] Poster is uploaded.
• [2024/12/17] 🔥 Scripts for evaluating RDT in Maniskill Simulation Benchmark is released!
• [2024/10/23] 🔥 RDT-170M (Smaller) model is released, a more VRAM-friendly solution 🚀💻.

Installation

1. Clone this repo and install prerequisites:

   # Clone this repo
   git clone [email protected]:thu-ml/RoboticsDiffusionTransformer.git
   cd RoboticsDiffusionTransformer
   
   # Create a Conda environment
   conda create -n rdt python=3.10.0
   conda activate rdt
   
   # Install pytorch
   # Look up https://pytorch.org/get-started/previous-versions/ with your cuda version for a correct command
   pip install torch==2.1.0 torchvision==0.16.0  --index-url https://download.pytorch.org/whl/cu121
   
   # Install packaging
   pip install packaging==24.0
   
   # Install flash-attn
   pip install flash-attn --no-build-isolation
   
   # Install other prequisites
   pip install -r requirements.txt
   

2. Download off-the-shelf multi-modal encoders:

   You can download the encoders from the following links:

   • t5-v1_1-xxl: link🤗
   • siglip: link🤗

   And link the encoders to the repo directory:

   # Under the root directory of this repo
   mkdir -p google
   
   # Link the downloaded encoders to this repo
   ln -s /path/to/t5-v1_1-xxl google/t5-v1_1-xxl
   ln -s /path/to/siglip-so400m-patch14-384 google/siglip-so400m-patch14-384
   

3. Fill the missing argument in this file:

   Note that this buffer will only be used during pre-training. See this doc for more details.

   # ...
   
   dataset:
   # ...
   # ADD YOUR buf_path: the path to the buffer (at least 400GB)
      buf_path: /path/to/buffer
   # ...
   

Fine-Tuning on Your Own Dataset

If your fine-tuning dataset is in the Open X-Embodiment or the collection of our pre-training datasets (see this doc), you can also fine-tune RDT through the pre-trained pipeline. You need to remove other redundant datasets in the parameters. We refer to this guide (pre-training).

1. Prepare your dataset:

   You need to download your dataset to the disk and give it a name my_cool_dataset.

   Then, you can link your dataset to the repo directory:

   # Under the root directory of this repo
   cd data
   mkdir -p datasets
   
   # Link the downloaded dataset to this repo
   ln -s /path/to/my_cool_dataset datasets/my_cool_dataset
   

2. Implement the dataset loader:

   You need to:

   1. Register the configuration of my_cool_dataset:

      Append the control frequency of my_cool_dataset in this file. Write the name of my_cool_dataset in this file and this file, where the value of the sampling weight doesn't matter since you only have one dataset. In these two files, we leave a placeholder of agilex; you can simply replace it with my_cool_dataset.

   2. Re-Implement the class of HDF5VLADataset:

      You can find this class in this file. In this file, we provide an example of loading the fine-tuning dataset used in our paper (see this link).

      To adapt it to your dataset, you need to: (a) modify the HDF5_DIR (directory to my_cool_dataset) and DATASET_NAME (should be "my_cool_dataset") in L21 and L22; (b) Implement the two functions of parse_hdf5_file() and parse_hdf5_file_state_only(). Please take a look at the original file for detailed comments and examples.

      Note 1: Despite its name, you don't necessarily need to use HDF5 to store your data. Just make sure that the class is correctly implemented.

      Note 2: During implementation, you may need to fill your robot action into the unified action vector (L180-194). Please refer to this file for an explanation of each element in the unified vector. We have reserved enough slots for each physical quantity. For example, we have reserved ten slots for joint angles. If your robot arm has six degrees of freedom, you only need to fill in the first six.

      IMPORTANT 1: If your robot is single-arm, please fill its action into the right-arm portion of the unified action vector, aligning with our pre-training datasets.

      IMPORTANT 2: We use 6D representation for EEF rotation. If your action space contains EEF rotation (angle or quaternion), please refer to this file for conversion. We note that this mapping is not reversible. Different Euler angles may be equivalent and correspond to the same 6D representation.

      IMPORTANT 3: No physical quantities (except the gripper width) are normalized during pre-training. This can preserve each physical quantity's meaning, thereby promoting generalization across robots. Therefore, we encourage you not to normalize any physical quantities but to choose appropriate units for them. Generally, we use the International System of Units, which ensures that most values fall within [-1,1]. As an exception, we perform min-max normalization on the gripper width to [0,1].

      IMPORTANT 4: If you use RTX 4090 (or lower), the GPU memory may be too low to load the t5-v1_1-xxl encoder. Instead, we recommend you precompute the language embeddings (see this file for an example script) and load them during training. In this way, you need to specify the path to the embeddings in the HDF5VLADataset (see L148) rather than the natural language.

   3. Compute the dataset statistics information for my_cool_dataset:

      # Under the root directory of this repo
      # Use -h to see the full usage
      python -m data.compute_dataset_stat_hdf5
      

3. Start fine-tuning:

   Configurations relevant to model architecture and data processing are in this file. Normally, you do not need to modify these configurations; otherwise, it will cause errors in loading the pre-training checkpoint. Configurations relevant to training are passed through Command Line Arguments. Use python main.py -h to see the descriptions. We provide an example of a fine-tuning script in this file (finetune.sh). You may need to modify some of the parameters in this file, such as CUTLASS_PATH and WANDB_PROJECT.

   Use this to start fine-tuning:

   source finetune.sh
   

   with finetune.sh detailed as below:

      deepspeed --hostfile=hostfile.txt main.py \
         --deepspeed="./configs/zero2.json" \   # If you want to use DeepSpeed, which is strongly recommended
         --pretrained_model_name_or_path=<MODEL ID | DIRECTORY OF MODEL WEIGHTS | PATH TO MODEL CHECKPOINT> \
         --pretrained_text_encoder_name_or_path=<MODEL ID | PATH TO MODEL DIRECTORY > \   # e.g., google/t5-v1_1-xxl
         --pretrained_vision_encoder_name_or_path=<MODEL ID | PATH TO MODEL DIRECTORY> \  # e.g., google/siglip-so400m-patch14-384
         --output_dir=<DIRECTORY to SAVE CHECKPOINTS> \ # e.g., checkpoints/rdt-1b-agilex
         --train_batch_size=32 \
         --sample_batch_size=64 \   # batch size for diffusion sampling in validation 
         --max_train_steps=200000 \
         --checkpointing_period=1000 \
         --sample_period=500 \   # sample period for validation
         --checkpoints_total_limit=40 \
         --lr_scheduler="constant" \
         --learning_rate=1e-4 \
         --mixed_precision="bf16" \ # If you want to use mixed precision, bf16 is recommended
         --dataloader_num_workers=8 \
         --image_aug \  # If you want to use image augmentation
         --dataset_type="finetune" \
         --state_noise_snr=40 \  # If you want to add noise to the state
         --load_from_hdf5 \   # If you use HDF5 to store your data
         --report_to=wandb
   

   IMPORTANT: If you have already chosen to precompute the language embeddings, please specify --precomp_lang_embed in the finetune.sh.

   Note 1: pretrained_model_name_or_path can one of:

   • a string, the model id of a pre-trained model hosted inside a model repo on HuggingFace. Please fill with "robotics-diffusion-transformer/rdt-1b", which is the officially-released RDT-1B model🤗 at HuggingFace. (recommended)
   • a string, the path to a directory containing the manually downloaded model weights from HuggingFace, e.g., "/path/to/rdt-1b". You should first manually download the rdt-1b directory from this link🤗.
   • a string, the path to a directory containing model weights saved using [~RDTRunner.save_pretrained] method. This can be either:
     • "checkpoints/rdt-pretrain-1b/checkpoint-<STEP NUMBER>": This is the path to the checkpoint saved in the <STEP NUMBE> iteration during pre-training. Refer to this file for a tutorial on how to start your own pre-training.
     • "checkpoints/rdt-pretrain-1b": If the pre-training completes normally without any exception, you can specify this path to load the last checkpoint.
   • a string, the path to model checkpoint (*.pt) saved by DeepSpeed, e.g., "checkpoints/rdt-pretrain-1b/checkpoint-<STEP NUMBER>/pytorch_model/mp_rank_00_model_states.pt" (verified)
   • None if you want to randomly initialize the model using configuration at config_path.

   Note 2: You can monitor the training process by observing loss (through a long window moving average) and overall_avg_sample_mse in Wandb or TensorBoard. We empirically found that the lower the overall_avg_sample_mse, the better the model performs. Usually, fine-tuning is over when this value converges.

   Note 3: If the training oscillates, you can increase the batch size by adding more GPUs or setting a larger --gradient_accumulation_steps.

   Note 4: Please specify --load_from_hdf5 in your script when finetuning with an HDF5 dataset.

Deployment on Real-Robots

We have encapsulated the inference of the model into a class named RoboticDiffusionTransformerModel (see this file). You can call this class's step() method for inference. However, you may need to re-implement some parts according to your specific robot. You should at least modify the _format_joint_to_state() (L164) and _unformat_action_to_joint() (L196) to convert between robot raw actions and unified action vectors that RDT accepts. You may also specify the control frequency of your robot (L49).

IMPORTANT: When you feed the images into step(), remember the order MUST be [ext_{t-1}, right_wrist_{t-1}, left_wrist_{t-1}, ext_{t}, right_wrist_{t}, left_wrist_{t}].

We provide an example hardware code in this file for deployment on Mobile ALOHA, and the corresponding running script in this file (inference.sh), which is detailed below;

   python -m scripts.agilex_inference \
      --use_actions_interpolation \
      --pretrained_model_name_or_path=<PATH TO MODEL CHECKPOINT> \  # your finetuned checkpoint: e.g., checkpoints/rdt-finetune-1b/checkpoint-<STEP NUMBER>, checkpoints/rdt-finetune-1b/checkpoint-<STEP NUMBER>/pytorch_model/mp_rank_00_model_states.pt, the same before
      --lang_embeddings_path=<PATH TO YOUR INSTURCTION EMBEDDINGS> \ # e.g. outs/lang_embeddings/your_instr.pt"
      --ctrl_freq=25    # your control frequency

IMPORTANT: If you on-board GPU memory is not enough to encode the language, please refer to this file for precomputation and specify the language embedding path in inference.sh. Detail instructions are provided below:

1. Set Required Parameters in scripts/encode_lang.py

   # ...
   
   GPU = 0
   MODEL_PATH = "google/t5-v1_1-xxl"
   CONFIG_PATH = "configs/base.yaml"
   SAVE_DIR = "outs/"   # output directory
   
   # Modify this to your task name and instruction
   TASK_NAME = "handover_pan"
   INSTRUCTION = "Pick up the black marker on the right and put it into the packaging box on the left."
   
   # Note: if your GPU VRAM is less than 24GB, 
   # it is recommended to enable offloading by specifying an offload directory. 
   OFFLOAD_DIR = None  # Specify your offload directory here, ensuring the directory exists.
   
   # ...
   

2. Run the script

   python -m scripts.encode_lang
   

Note: If you want to deploy on the Mobile ALOHA robot, don't forget to install the hardware prerequisites (see this repo).

Simulation Benchmark

We comprehensively evaluate RDT against baseline methods using the ManiSkill simulation benchmark. Specifically, we focus on five benchmark tasks: PegInsertionSide, PickCube, StackCube, PlugCharger, and PushCube. Here's a brief overview of the evaluation setup:

Evaluation Setup:

1. Install ManiSkill:
   Within the RDT environment, install ManiSkill as follows:

   conda activate rdt
   pip install --upgrade mani_skill
   

2. Configure Vulkan:
   Follow the ManiSkill documentation to properly set up Vulkan。

3. Obtain Model Weights:
   Download the fine-tuned model weights from this Hugging Face repository. Download the precomputed language embeddings from here to the root directory of this repo.

4. Run Evaluation Scripts:
   After completing the setup steps, execute the provided evaluation scripts to assess RDT on the selected tasks.

conda activate rdt 
python -m eval_sim.eval_rdt_maniskill \
--pretrained_path PATH_TO_PRETRAINED_MODEL

Implementation Details

Data

Utilizing the official ManiSkill repository, we generated 5,000 trajectories through motion planning. The initial action mode of these trajectories is absolute joint position control and we subsequently converted them into delta end-effector pose control to align with the pre-training action space of OpenVLA and Octo. We strictly adhered to the official codebases of OpenVLA and Octo, modifying only the dataset-loading scripts. Consequently, we finetuned OpenVLA and Octo using the delta end-effector pose data. For RDT and Diffusion-Policy we leverage joint position control data for training which is aligned with our pre-training stage as well.

Training

• OpenVLA is fine-tuned from the officially released pre-trained checkpoint with LoRA-rank 32 until converge.
• Octo is fine-tuned from the officially released pre-trained checkpoint for 1M iterations until converge.
• Diffusion-Policy is trained from scratch for 1000 epochs. We select the checkpoint of 700 epoch which has the lowest validation sample loss of 1e-3.
• RDT is fine-tuned from our released pre-trained checkpoint for 300ks iterations.

Results

Each method is evaluated over 250 trials (10 random seeds with 25 trials per seed). The quantitative results, including success rate mean and std value across 10 random seeds are presented below:

Finetune RDT with Maniskill Data

To fine-tune RDT with Maniskill data, first download the Maniskill data from here and extract it to data/datasets/rdt-ft-data. Then copy the code in data/hdf5_vla_dataset.py to data/hdf5_maniskill_dataset.py and run the following script:

bash finetune_maniskill.sh

Reproducing Baseline Results

Download and extract the fine-tuned model weights from here to eval_sim/.

• OpenVLA: Clone OpenVLA repo in ./eval_sim/ and install its environment & ManiSkill. Then run the following script:

python -m eval_sim.eval_openvla --pretrained_path PATH_TO_PRETRAINED_MODEL

• Octo: Clone Octo repo in ./eval_sim/ and install its environment & ManiSkill. The run the following script:

python -m eval_sim.eval_octo --pretrained_path PATH_TO_PRETRAINED_MODEL

• Diffusion-Policy: Clone our simplified Diffusion-Policy repo in ./eval_sim/ and run:

python -m eval_sim.eval_dp --pretrained_path PATH_TO_PRETRAINED_MODEL

RDT on RoboTwin Benchmark

RoboTwin is a benchmark simulator based on Sapien, comprising 50 common dual-arm tasks. All policies are trained on 50 trajectories collected in a clean environment for each of the 50 tasks, and then deployed and tested for success rates in environments corresponding to either easy (clean) or hard (randomized) configurations for the same 50 tasks. You can refer to ​​https://robotwin-platform.github.io/doc/usage/RDT.html#1-environment-setup​​ to set up the environment required for RoboTwin and RDT. According to the official test results of RoboTwin (RoboTwin 2.0 Leaderboard), RDT ranks second only to Pi0 (excluding the DP3 policy, which utilizes ground truth point clouds).

FAQ

1. How can I fine-tune RDTs with limited VRAM?

• Use a Smaller Model: Opt for the RDT-170M model, which requires less VRAM.

• Select a Memory-Efficient ZeRO Stage: Choose a more memory-efficient ZeRO stage based on your needs:

  • ZeRO-3 with Offload > ZeRO-3 > ZeRO-2 with Offload > ZeRO-2 > ZeRO-1
  • By default, we use ZeRO-2 for a balance between speed and memory efficiency. Find more details on ZeRO stages here and here.

• Enable 8-bit Adam Optimization: Activate 8-bit Adam by setting use_8bit_adam=True for reduced memory usage during training.

• Apply 4-bit or 8-bit Quantization: Quantizing model weights can significantly reduce VRAM requirements.

• Use XFormers: This library provides optimized transformers with efficient memory usage.

• Enable Gradient Checkpointing: Implement gradient_checkpointing manually to save memory during backpropagation. See here for instructions. Once you have successfully implemented this feature, we welcome you to submit a PR👏.

• Gradient Accumulation: Set a larger --gradient_accumulation_steps=<num_steps>. This will accumulate the gradients of <num_steps> batches for backpropagation. Equivalently, this will increase the batch size by <num_steps> times, at the cost of <num_steps> times the running time.

2. How many steps are recommended for fine-tuning RDT?

Regardless of the batch size you select, it is recommended to train for at least 150K steps to achieve optimal results.

3. What to do if t5-xxL is too large to store in GPU memory?

1. Do not load T5-XXL in your GPU memory when training. Pre-compute language embeddings in advance.
2. Set OFFLOAD_DIR to enable CPU offloading in scripts/encode_lang_batch.py and scripts/encode_lang.py.
3. Use smaller versions of t5 like t5-base instead of t5-xxL.

Citation

If you find our work helpful, please cite us:

@article{liu2024rdt,
  title={RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation},
  author={Liu, Songming and Wu, Lingxuan and Li, Bangguo and Tan, Hengkai and Chen, Huayu and Wang, Zhengyi and Xu, Ke and Su, Hang and Zhu, Jun},
  journal={arXiv preprint arXiv:2410.07864},
  year={2024}
}

Thank you!

License

All the code, model weights, and data are licensed under MIT license.