arXiv | Project Website
This repository contains the DynaGuide implementation code. This includes code to train the dynamics model and use the model during inference-time to steer a diffusion policy. We include example uses of DynaGuide as seen in our paper experiments: a toy block environment and the Calvin environment. Finally, we include code snippets that can be added to any diffusion policy codebase to enable DynaGuide.
- 6/16/25: Uploaded all code and example of pretrained dynamics model, base policy, and guidance condition.
- Install pytorch
- Go to
robomimic/and runpip install -e . - Go to
calvin/and runinstall.sh - Go to this directory and run
pip install -e . - Download the calvin datasets and process them according to this section. Or, for the toy experiments, collect the data according to this section. For your own dataset / diffusion policy, refer to this section for general deployment of DynaGuide.
The core of DynaGuide is found in the core/ folder in this directory. Experiment configs are found in configs/ (base policy training) and calvin_exp_configs_examples/ (DynaGuide evals). Auxiliary data processing scripts are found in data_processing_calvin/. See "Generating the Dataset Splits" for instructions on using them. Scripts for getting the paper results are found in paper_experiments/, and the toy experiment can be found in toy_squares_experiment. Figure generation can be found in figure_generation/ that takes raw experiment output and creates the figures in the paper.
The modified Calvin environment can be installed from the Calvin/ folder, and the diffusion policy can be installed from the robomimic/ folder. For more details, refer to the next section.
The experiment code can be found in the main directory here
analyze_calvin_touch.py: the code that looks at the DynaGuide results and gives statistics on individual behaviors; this code is repeated with modifications in the figure generation folder.run_dynaguide.py: the code that evaluates dynaguide and records the data for analysis. It also runs the base policy directly and goal conditioned policies.run_sampling_baseline.py: the code that runs the baseline where the base policy is sampled multiple times and the best action is keptrun_trained_agent.py: not used often, but this is the default code for evaluating trained base policiestest_dynaguide_embedding.py: the code that takes a trained dynamics model and runs a suite of tests to ensure that is functional before applying DynaGuidetrain_base_policy.py: this is the code that trains the base diffusion policytrain_dynaguide.py: this is the code that trains the dynamics model
Additionally, the diffusion policy in robomimic/robomimic/algo/diffusion_policy.py is modified to support guidance. You can modify any DDIM noise prediction diffusion policy to support guidance (see this section)
This codebase contains the DynaGuide method, and we test it on the Calvin benchmark environment.
You will find the scripts for processing CALVIN data in the data_processing_calvin folder. You will run these scripts as follows:
- Download Calvin dataset (train and validation) using their provided instructions. Use the ABCD split for dynamics model training, and the D split for the base policy training.
- Run
calvin_to_labeled_hdf5.pyto segment the behaviors into a .h5 file. Do it once for the Calvin train split, and this will be your training .h5 file you use intrain_dynaguide.py. Do it again for the Calvin validation split, and this will be used for the guidance conditions and the tests for the dynamics model. - For the validation set: run
split_behavioral_validation_datasets_calvin.pyto separate these segmented behaviors to individual h5 files and a mixed test file for testing the dynamics model. - To create the guidance conditions, use these individual h5 files in
hdf5_combiner.py, which can create single h5 files with multiple behaviors with a set number of conditions per behavior
Additional notes
- To remove individual trajectories from guidance conditions files (e.g. there's a bad segmentation), use
hdf5_filter.py. - To extract initial robot states for use in experiment resets (found in the config file), use
retrieve_initial_robot_states.py - To extract goal images for the goal conditioning baseline, use
generate_goal_images.py
At this point, you should have
- An h5 dataset for base policy training, downloaded & parsed from the Calvin D train split
- An h5 dataset for dynamics model training, downloaded & parsed from the Calvin ABCD train split.
- h5 datasets for dynamics model testing, including a multi-behavior test h5 file and individual behavior h5 files, created from the
split_behavioral_validation_datsets_calvin.py
Base Policy
The base policy is a diffusion policy implemented in Robomimic. We can use the provided robomimic script train.py to train the base policy. We provide the configs in the configs/ folder.
python split_train_val.py --dataset path_to_calvin_dataset --ratio 0.03
python train.py --config configs/diffusion_policy_image_calvin_ddim.json \
--dataset path_to_calvin_dataset --output output_folder --name run_name
Dynamics Model The dynamics model is trained using this DynaGuide codebase. For the calvin environment, you can train the model with this configuration:
python train_dynaguide.py --exp_dir directory_to_save_dynamics_model \
--train_hdf5 your_calvin_dynamics_dataset \
--test_hdf5 your_calvin_validation_dataset \
--cameras third_person --action_dim 7 --proprio_key proprio --proprio_dim 15 \
--num_epochs 6000 --action_chunk_length 16 --batch_size 16
To test the model, test_dynaguide.py contains a set of visualizations to analyze the ability for the dynamics model to predict the future and take actions into account. To run these tests, use this configuation:
python test_dynaguide_embedding.py --exp_dir directory_of_dynamics_model \
--good_hdf5 single_target_behavior_h5_file \
--mixed_hdf5 mixed_behavior_test_h5_file \
--checkpoint path_to_dynamics_model_checkpoint \
--action_chunk_length 16 --key button_off
The DynaGuide experiments in the Calvin environment require an experiment config that dictates the environment setup, guidance conditions, and others. Examples of these are included in calvin_exp_configs_examples, including the ones used in the paper experiments.
{
"env_setup" : {"green_light" : 1}, -> This is how to set up the environment for a particular test. See below for more details
"use_neg" : false, -> make true if you're including negative guidance conditions
"pos_examples": "/yourfolderhere/dataset/CalvinDD_validation_by_category_wcubes/button_off_20.hdf5", -> path to positive guidance conditions.
"loc_target": [ -> For the position guidance baseline only
-0.11221751715334842,
-0.11465078614523619,
0.4966200809334477
],
"reset_poses": "initial_calvin_robot_states_midpoint.json" -> file containing reset poses for the robot (provided)
}
The env_setup tells the environment how to configure the articulated elements. This is important for testing the ability for DynaGuide to guide towards one particular behavior.
green_light{0, 1} sets the green light to either on or offsliding_door[0, 0.28] sets the sliding door to a position. The base posiion (0) is the door fully right. At 0.28, the door is left.drawer[0, 0.16] sets the drawer to a position. T he base position (0) is the drawer closed. At 0.16, the drawer is open.switchsets the lever switch to a position (and the light is set automatically). The base position (0) is the switch in the off position. At 0.085, the switch is in the on position.
To run DynaGuide, you need 1) the experiment config file 2) trained base policy 3) trained dynamics model and 4) guidance conditions (h5 files). Once you have these components, you can run run_dynaguide.py using the following code:
python run_dynaguide.py --video_path $output_folder/$run_name.mp4 \
--dataset_path output_hdf5_path --dataset_obs --json_path output_json_path --horizon 400 --n_rollouts 100 \
--agent checkpoint_dir --output_folder output_folder --video_skip 2 \
--exp_setup_config exp_config_path --guidance dynamics_model_path \
--camera_names third_person --scale 1.5 --ss 4 --alpha 30 --save_frames
For examples of this evaluation call, look at the shell files in paper_experiments. Note that the scale, ss, and alpha can be modified here.
For your convenience, the scripts used to run the paper experiments are included in paper_experiments/. Modifications will be needed for different system capacities and paths.
- Experiment 1:
calvin_articulated_object.shand generate plots withexperiment_1_graphs.py - Experiment 2:
calvin_movable_object.shand generate plots withexperiment_2_blocks.py - Experiment 3:
calvin_underspecified_objectives.shand generate plots withexperiment_3_partialgoals_graphs.py - Experiment 4:
calvin_multiple_behaviors.shand generate plots withexperiment_4_multiobjective.py - Experiment 5:
calvin_underrepresented_behaviors.shand generate plots withexperiment_5_deprivation_graphs.py
- Download the DynaGuide model from this link
- Download the Base policy from this link
- Download the Guidance conditions for SWITCH_ON from this link
- Modify the
calvin_exp_configs_examples/switch_on.jsonby changing thepos_examplesfile path to the downloaded guidance condition - Make a folder called
resultsin this directory Run the following code:
run_name=SwitchOnDynaGuide
output_folder=results/$run_name
checkpoint_dir=path_to_base_policy
exp_setup_config=calvin_exp_configs_examples/switch_on.json
embedder=path_to_embedder
python run_dynaguide.py --video_path $output_folder/$run_name.mp4 \
--dataset_path $output_folder/$run_name.hdf5 --dataset_obs --json_path $output_folder/$run_name.json --horizon 400 --n_rollouts 100 \
--agent $checkpoint_dir --output_folder $output_folder --video_skip 2 \
--exp_setup_config $exp_setup_config --guidance $embedder --camera_names third_person --scale 1.5 --ss 4 --alpha 30 --save_frames
To compare with base policy, run the control using this code
run_name=BasePolicy
output_folder=results/$run_name
checkpoint_dir=path_to_base_policy
exp_setup_config=calvin_exp_configs_examples/switch_off.json
embedder=path_to_embedder
python run_dynaguide.py --video_path $output_folder/$run_name.mp4 \
--dataset_path $output_folder/$run_name.hdf5 --dataset_obs --json_path $output_folder/$run_name.json --horizon 400 --n_rollouts 100 \
--agent $checkpoint_dir --output_folder $output_folder --video_skip 2 \
--exp_setup_config $exp_setup_config --guidance $embedder --camera_names third_person --scale 0 --ss 1 --save_frames
Finally, fill the experiment name and directory in analyze_calvin_touch.py and run the code to see the behavior distribution before and after DynaGuide
See above instsructions for how to run the experiments seen in the DynaGuide paper.
The toy square touching experiment is found in toy_squares_experiment/ and the environment is found in robomimic/robomimic/envs/env_flat_cube.py. Unlike the main DynaGuide method, this dynamics model is actually a classifier and it is trained end-to-end with synthetic data. To collect this data, run this code:
run_name=pymunk_touch_res128_largercubes_repeated_100ktrain
output_folder=output_folder_here
python collect_scripted_data_pymunk.py --video_path $output_folder/$run_name.mp4 \
--dataset_path $output_folder/data.hdf5 --dataset_obs --json_path $output_folder/config.json --horizon 150 --n_rollouts 100000 \
--env_config /store/real/maxjdu/r
8F09
epos/robotrainer/configs/touchcubes.json --output_folder $output_folder --keep_only_successful \
--camera_names image --video_skip 5 --repeat_environment
Then you can train the dynamics model using this code:
experiment_name=Pymunk_classifier_FROMSCRATCH_100k_noised_ddim
CUDA_VISIBLE_DEVICES=6 python train_end_state_classifier.py --exp_dir /store/real/maxjdu/repos/robotrainer/results/classifiers/$experiment_name/ \
--train_hdf5 /store/real/maxjdu/repos/robotrainer/dataset/pymunktouch/pymunk_touch_res128_largercubes_repeated_100ktrain/data.hdf5 \
--test_hdf5 /store/real/maxjdu/repos/robotrainer/dataset/pymunktouch/pymunk_touch_res128_largercubes_repeated_valid/data.hdf5 \
--num_epochs 12000 --action_chunk_length 16 --batch_size 16 --noised
and run tests with this code:
experiment_name=Pymunk_classifier_FROMSCRATCH_100k_noised_ddim
checkpoint=11900
CUDA_VISIBLE_DEVICES=6 python test_end_state_classifier.py --exp_dir /store/real/maxjdu/repos/robotrainer/results/classifiers/$experiment_name/ \
--mixed_hdf5 /store/real/maxjdu/repos/robotrainer/dataset/pymunktouch/pymunk_touch_res128_largercubes_repeated_valid/data.hdf5 \
--checkpoint /store/real/maxjdu/repos/robotrainer/results/classifiers/$experiment_name/$checkpoint.pth \
--action_chunk_length 16
Finally, you can run an experiment by running code like this:
CUDA_VISIBLE_DEVICES=5 python classifier_guidance_pymunk.py --video_path output_video_path \
--dataset_path output_data_path --dataset_obs --json_path output_json --horizon 400 --n_rollouts 100 \
--agent base_policy_path --output_folder output_folder_path --video_skip 1 \
--guidance classifier_ckpt_path --scale 0.5 --camera_names image --target_list 1,-0.33,-0.33,-0.33 --render_visuals
The target_list allows you to specify which colors to seek and avoid. Positive means seek, and negative means avoid.
The code found in core/ should be general and applicable to any diffusion-based policy. To deploy it on any such policy, you can train the dynamics model using the existing DynaGuide scripts and compute the guidance function using the existing code in core/dynaguide.py. To deploy it to any diffusion policy:
- Locate the function that does the inference-time denoising process
- Verify that the model is indeed predicting noise
- Add this code:
scaled_grad = guidance_function(state, naction)
noise_pred = noise_pred - (1 - noise_scheduler.alphas_cumprod[k]).sqrt() * scaled_grad
where guidance_function is the function computed by core/dynaguide.py and noise_pred is the predicted noise outputted by the noise estimation network. Use the noise scheduler of the diffusion policy as noise_scheduler. If you use diffusers.schedulers.scheduling_ddim, nothing needs to be changed.
Add this after the noise_pred is updated by the network and before the noise scheduler operates. To enable stochastic sampling, simply wrap the denoising step in an inner for-loop. An example is shown here:
for ss_iter in range(ss): # stochastic sampling; notice how the "k" doesn't decrease in this loop.
noise_pred = nets['policy']['noise_pred_net'](
sample=naction,
timestep=k,
global_cond=obs_cond
)
state = inputs["obs"]
scaled_grad = guidance_function(state, naction)
noise_pred = noise_pred - (1 - self.noise_scheduler.alphas_cumprod[k]).sqrt() * scaled_grad
# inverse diffusion step (remove noise)
naction = self.noise_scheduler.step(
model_output=noise_pred,
timestep=k,
sample=naction
).prev_sample # this gives the mu from the paper
@misc{du2025dynaguidesteeringdiffusionpolices,
title={DynaGuide: Steering Diffusion Polices with Active Dynamic Guidance},
author={Maximilian Du and Shuran Song},
year={2025},
eprint={2506.13922},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2506.13922},
}
We are grateful for the Calvin benchmark codebase duplicated here in this repository with minor modifications. The diffusion policy is adapted from the Robomimic codebase also duplicated here with modifications. The dynamics model was inspired by the Dino-WM implementation and leverages visuals representations from Dino-V2.
Are you trying to get DynaGuide working and something's wrong? The numbers don't look right? Send me an email at maxjdu@stanford.edu !