DexLab#

What are we talking about
when we talk about
dexterous simulation?

From contact and touch to policies, evaluation and real-world transfer. A guide to the papers, projects and evidence behind dexterous simulation.

Read the research guide

The research guide is in Chinese.

Agents and simulation, improving together.

Try a taskDiagnose failureImprove skills · Improve simulationIndependently retest↩ Ask the next question

I study how agents can help improve manipulation skills and physics simulation through experiments: humans set goals and boundaries; agents help identify questions, experiment and iterate. A complete autonomous loop remains unverified; current results come from individual experiments.

Research findings

More research findings
Recorded replay · Left: point contact / right: synthetic torsion · Same time, original speed · Click to play

Adding torsional contact passed the hold check.

  1. Finding

    With the same controller, synthetic torsional contact reduces maximum two-second drift from 2.734 to 1.025 mm, passing the original 2 mm limit and active-release checks.

  2. Evidence

    Both original point contacts persist but lack torque about their connecting line. Offset gravity drives rotation and the hold fails; more normal force does not add that missing direction.

  3. Takeaway

    For this kind of grasp failure, check whether the contact model can supply the required torque before discussing parameters or solving. This comparison uses existing MuJoCo capability; the solver is unchanged.

  4. Scope

    The finding is limited to this synthetic contact and two-second window; slow drift remains. The new law also affects preparation, so not all displacement can be uniquely attributed to one degree of freedom.

Read the analysis and evidence
What does this result establish?

The earliest identified issue is a structural gap in the contact wrench; offset gravity drives rotation about the contact line. The line is intuitively like an axis without torsional resistance, but moving fingers and contact points mean it is not a fixed hinge. The new law also acts during preparation, changing the initial state; this development comparison does not uniquely attribute all displacement to one degree of freedom.

The result is limited to the synthetic sphere, this controller and the two-second window; slow drift remains. Longer holds, contact switching and new validation conditions are the next tests. Real fingertip calibration and a complete autonomous improvement process remain unverified.

Follow a research question

Start with the research guide (中文)

Why can a tighter grasp still slip?

Start with the forces and torques a task needs, then examine models, parameters and solving.

Connect to the learning map
  1. Point versus torsional contact: a grasp comparison
  2. Does lower prediction error reveal contact failures?
  3. What belongs to the model, numerical problem or solver?

When does task success mean better capability?

Trace actions, state and forces to the first failing layer. Read task scores alongside physical evidence.

Connect to the learning map
  1. Failure atlas: finding the first failing layer
  2. The gripper is closed. Why does the bread fall?
  3. Revisiting wiping without assisted retention

How can one failure inform the next improvement?

Connect skills, observations, simulation and independent evaluation to choose the next test. A complete autonomous improvement process remains a research direction.

Connect to the learning map
  1. RoboDojo / Astra: corrections and reliable contact control
  2. Sharing policies and dynamics across demonstrations
  3. Separating the model, execution and scoring

Simulation learning

Start with MuJoCo: learn time, contact and solving through an interactive example and self-checks, then explore the detailed courses and version 4.

Start learning MuJoCo
  1. 1

    Build physical intuition

    State, time & dynamics

    Start: vectors, matrices & basic mechanicsPhysics-Based Simulation study route
  2. 2

    Understand contact & friction

    Geometry, forces, torques & constraints

    Prerequisite: physical and state foundationsContact lesson
  3. 3

    Read the numerical problem

    Equations, iteration, steps & tolerances

    Prerequisite: contact models & linear algebraNine-part solver guide
  4. 4

    Explore a native engine

    Use the API and trace the source

    Continue along the application or theory routeChoose an engine learning path
Choose an engine to explore

Go deeper through the related concepts.

Sim Learning courses are primarily in Chinese, with English terminology and reading summaries.