Hi @jake-levy and @tyler-westenbroek,
First, thank you for open-sourcing the Simulation Distillation (SimDist) framework and the accompanying paper. The work on distilling structural priors from simulation into a latent world model for rapid real-world adaptation is very inspiring, and I've been working through the codebase to reproduce and build on your results.
I'm writing because I'm currently working on extending SimDist for a research project, and I've hit a practical bottleneck that I believe your group could help with.
Background
I have access to compute for training but, unfortunately, my lab does not currently have a Unitree Go2 (or comparable quadruped) available for data collection. The project documentation (docs/deployment_go2.md and docs/adaptation.md) clearly indicates that real-world fine-tuning relies on trajectories logged via the go2_ros2_ws stack — typically stored under datasets/real/.
I noticed that the repository currently only ships simulation datasets (datasets/sim/); the datasets/real/ directory is intentionally gitignored, which makes sense for privacy/storage reasons, but it does mean there is no public way to test the "Real-World Go2" → "Adaptation" pipeline without hardware.
Request
Would it be possible for your group to share (or point me to) the real-world trajectories that were used in the experiments reported in the RSS2026 paper? Specifically, I'm hoping to get something along the lines of:
Raw HDF5 episodes (the format written by host_full_record.sh / auto_record.sh), or
An aggregated/processed dataset in the layout produced by aggregate_realworld_data.py + process_data.py (under datasets/real//processed_data_*),
so that I can run:
python scripts/finetune_model.py
data.dataset_name=<real_dataset>
checkpoint.resume_checkpoint=<pretrained_model>
run_name=<my_run_name>
and evaluate the sim-to-real adaptation pipeline on the same data you used.
Hi @jake-levy and @tyler-westenbroek,
First, thank you for open-sourcing the Simulation Distillation (SimDist) framework and the accompanying paper. The work on distilling structural priors from simulation into a latent world model for rapid real-world adaptation is very inspiring, and I've been working through the codebase to reproduce and build on your results.
I'm writing because I'm currently working on extending SimDist for a research project, and I've hit a practical bottleneck that I believe your group could help with.
Background
I have access to compute for training but, unfortunately, my lab does not currently have a Unitree Go2 (or comparable quadruped) available for data collection. The project documentation (docs/deployment_go2.md and docs/adaptation.md) clearly indicates that real-world fine-tuning relies on trajectories logged via the go2_ros2_ws stack — typically stored under datasets/real/.
I noticed that the repository currently only ships simulation datasets (datasets/sim/); the datasets/real/ directory is intentionally gitignored, which makes sense for privacy/storage reasons, but it does mean there is no public way to test the "Real-World Go2" → "Adaptation" pipeline without hardware.
Request
Would it be possible for your group to share (or point me to) the real-world trajectories that were used in the experiments reported in the RSS2026 paper? Specifically, I'm hoping to get something along the lines of:
Raw HDF5 episodes (the format written by host_full_record.sh / auto_record.sh), or
An aggregated/processed dataset in the layout produced by aggregate_realworld_data.py + process_data.py (under datasets/real//processed_data_*),
so that I can run:
python scripts/finetune_model.py
data.dataset_name=<real_dataset>
checkpoint.resume_checkpoint=<pretrained_model>
run_name=<my_run_name>
and evaluate the sim-to-real adaptation pipeline on the same data you used.