DexSim Unified Coupled Solver For Embodied AI

DexSim Unified Coupled Solver For Embodied AI#

Developed by the DexForce Simulation Engine Team, DexUni is the recommended Newton solver configuration for EmbodiChain scenes that combine Articulated Objects with cloth or a volume soft body. DexSim exposes it through DexUniSolverCfg; EmbodiChain also accepts the short mapping name solver_type="dexuni".

DexUni is provided through the DexSim integration rather than Newton’s public native solver list. This page focuses on its user-facing capabilities. For Newton-native solver details, use the official Newton Solver Guide.

Capabilities#

Capability

Current support

Articulated Objects with cloth

Supported

Articulated Objects with volume soft bodies

Supported

Contact between Articulated Objects and deformables

Supported

Batched arenas, reset, state access, and rendering

Supported through SimulationManager

Fluid

Unsupported in the current version

Cable

Unsupported in the current version

Use DexUni for tasks in which Articulated Objects manipulate cloth, garments, cushions, or other supported deformable assets. If a scene has only deformables or only rigid bodies, let AutoSolver select the simpler matching solver unless the task needs an explicit choice.

The current version does not support fluid or cable simulation.

Minimal configuration#

AutoSolver selects DexUni when the complete scene contains both an articulation and a supported deformable. To keep the choice explicit:

from embodichain.lab.sim.cfg import NewtonPhysicsCfg

physics_cfg = NewtonPhysicsCfg(
    device="cuda:0",
    solver_cfg={"solver_type": "dexuni"},
    collision_cfg=None,
)

Add all Articulated Objects and deformables before SimulationManager.prepare(). DexUni requires GPU execution and requires_grad=False. Set collision_cfg=None because this solver owns the robot/deformable contact path it uses.

For prescribed motion, register the complete joint trajectory before prepare():

sim.add_robot(robot_cfg)
sim.add_deformable_object(deformable_cfg)
sim.register_kinematic_joint_trajectory("robot", joint_positions)
sim.prepare()

The trajectory layout is [num_envs, frames, dof], in the articulation’s public joint order.

Current limitations#

Warning

The current DexUni path does not solve or publish rigid-rigid or rigid-ground contacts. ContactSensor is therefore unsupported on this path.

Full capability coverage, including fluid, cable, and the rigid-contact limitations above, will be completed in the next release.

Qualify any task that mixes deformables with loose rigid objects against its required contact pairs. See Newton Runtime Qualification for the complete solver/device matrix and validation workflow.

Examples and further reading#

The examples are scene-specific starting points rather than universal tuning defaults. For deformable asset configuration and runtime data access, see Soft Object and Cloth Object.

Contents