Skip to content
huangkikiPublic

About

A Benchmark and Platform for Contact-Rich Manipulation

Resources

Stars

24 stars

Watchers

1 watching

Forks

Repository files navigation

DexLab

Contact-rich robot manipulation, beyond rigid-body simulation.

Research overview · Experiments · Sim-to-Real Gap · Learn simulation

DexLab is my personal research project exploring how Agent Skills and simulation can improve each other.

A grasp can fail because of the policy or because the simulator does not handle contact and friction accurately.

DexLab starts with concrete problems like this. Use tasks to expose problems, run controlled experiments to identify their causes, then improve the relevant part of the simulation.

The improved simulation helps the Agent learn new Skills. With those Skills, the Agent can take on harder tasks and probe the limits of the simulation.

DexLab architecture: tasks and protocols, systems under test, action execution, environment backends, and evidence and scoring
DexLab workflow: tasks and protocols → systems under test → action execution → environment backends → independent evidence and scoring.

Selected experiments

  • Contact model changes the verdict: point versus torsional contact in a two-finger sphere grasp.
  • Cutting is checked beyond the animation: tofu and banana records verify piece connectivity, contact impulse and independent motion.
  • A grasp is not a fold: cloth experiments score deformation, penetration, residual motion and release separately.
  • Engines are compared per profile: six-engine incline results retain failures and validity checks instead of producing a ranking.

Browse all experiments →

Evidence boundary

There are currently no hardware experiments. “Physical correctness” here means reproducible consistency inside a frozen simulation model: auditable inputs, analytic or engineering references, positive and negative controls, conservation checks and regression after each declared parameter change. It does not claim real-material calibration or sim-to-real accuracy.

Read the validation boundary →

Citation

@software{dexlab2026,
  author  = {Kiki Huang},
  title   = {DexLab: A Physics-Centric Research Platform for Contact-Rich Robot Manipulation},
  year    = {2026},
  url     = {https://github.com/huangkiki/Dexlab}
}

Email · GitHub

About

A Benchmark and Platform for Contact-Rich Manipulation

Resources

Stars

24 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages