Tactile Genesis

Note: this is not the final version of the paper. We plan to update all experiments with the latest version of Genesis World.

Self-contained code snapshot reproducing Tactile Genesis: Simulating Tactile Sensors at Scale to Explore Dexterous Task Learning, with the project page at neuroagents-lab.github.io/tactile-genesis.

If you want tactile sensors for your own project, use the latest Genesis World physics engine, which includes our tactile sensors, and optionally Eden (not yet available, but to be released before the end of 2026) if you need a managed learning framework.

Structure

Directory	Package	Role
dexterous-hands/	tactile_genesis	Tactile sensors, the three studied hand-task pairs, the distillation sweep, and every figure pipeline.
Eden/	eden	RL environment framework built on Genesis + rsl_rl. Declarative Pydantic configs, manager/term registries.
Genesis/	genesis-world	The physics-simulation backend (rigid/MPM/SPH/FEM/PBD) and the tactile sensor implementations. Upstream v1.4.1 (a68b97b9), the revision the paper's runs pin — see Genesis/VENDORED.md.
rsl_rl/	rsl-rl-lib	PPO + student/teacher distillation training loop.
docs/	—	The project page, served by GitHub Pages.

Dependency topology: dexterous-hands → Eden → Genesis, with rsl_rl as the training loop.

How the packages are wired

dexterous-hands/pyproject.toml resolves the three dependencies from the sibling checkouts via [tool.uv.sources] (editable path installs). These sources override the package origin across the whole resolution graph, so Eden's own transitive genesis-world / rsl-rl-lib requirements also point at the local checkouts:

[tool.uv.sources]
eden          = { path = "../Eden",    editable = true }
genesis-world = { path = "../Genesis", editable = true }
rsl-rl-lib    = { path = "../rsl_rl",  editable = true }

Quickstart

Everything is run from dexterous-hands/:

cd dexterous-hands
uv sync  # installs eden, genesis-world, rsl-rl-lib from local
uv run python main.py --task=in_hand_repose --robot=wuji --stage=1 --cpu \
  --config=conf/experiments/tiny.yaml --no_loaded_video --no_checkpoint_video

See dexterous-hands/README.md for the sensor interface, the distillation sweep, the figure pipelines, and what is not reproducible from this snapshot.

Project page

docs/ is a static site with no build step. Its three interactive pieces — the throughput benchmark, the prior-simulator comparison and the hand-layout browser — read JSON under docs/static/data/, regenerated from the committed benchmark logs and sensor layouts by scripts/build_site_data.py (which also copies the placement previews into docs/static/images/layouts/). The distillation curves are injected into the page itself:

python scripts/build_site_data.py
python dexterous-hands/scripts/build_ablation_widget.py \
  dexterous-hands/docs/experiments/distill-results-2999.json docs/index.html \
  --teachers dexterous-hands/docs/experiments/teacher-scores-eval17.json --final-iteration 2500

Preview it with python -m http.server -d docs 8000.

Licence

Apache 2.0 (see LICENSE). Redistributed robot models keep their own licences, listed in NOTICE — note that the XELA V4 tactile model under dexterous-hands/src/assets/hands/allegro/xela_v4/ is GPL-3.0.

Citation

If you use or reference our simulated tactile sensors in any way, please cite:

@article{chung2026tactilegenesis,
  title   = {Tactile Genesis: Simulating Tactile Sensors at Scale to Explore Dexterous Task Learning},
  author  = {Chung, Trinity and Yamazaki, Kashu and Patel, Dhruv and Duburcq, Alexis and Qiao, Yiling and Fragkiadaki, Katerina and Nayebi, Aran},
  journal = {arXiv preprint arXiv:2606.22332},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.22332}
}