Newton Physics Backend

Newton Physics Backend#

The Newton backend integrates Newton through DexSim while preserving EmbodiChain’s shared scene, object, sensor, and stepping APIs. Select it with NewtonPhysicsCfg, or with physics: newton in a Gym environment configuration.

Important

For scenes that combine an articulated robot with cloth or a volume soft body, start with DexUni. It is the DexSim coupled solver configuration for this mixed scene class and is the main Newton path documented by EmbodiChain.

DexSim Unified Coupled Solver For Embodied AI#

DexUni supports scenes that combine articulated robots with cloth or volume soft bodies. It is selected automatically when a finalized scene contains both an articulation and a supported deformable, or explicitly with solver_type="dexuni".

from embodichain.lab.sim import SimulationManagerCfg
from embodichain.lab.sim.cfg import NewtonPhysicsCfg

sim_config = SimulationManagerCfg(
    physics_cfg=NewtonPhysicsCfg(
        device="cuda:0",
        physics_dt=0.01,
        num_substeps=10,
        solver_cfg={
            "solver_type": "dexuni",
            "iterations": 8,
        },
        # DexUni owns particle-shape collision generation.
        collision_cfg=None,
    )
)

DexUni is a DexSim integration configuration, not another native solver class in Newton’s public solver list. Its capabilities and current EmbodiChain limitations are documented in DexSim Unified Coupled Solver For Embodied AI.

Automatic solver selection#

Leaving solver_cfg=None preserves DexSim’s scene-aware AutoSolverCfg. DexSim resolves it when SimulationManager.prepare() finalizes the complete Spawn scene:

Finalized EmbodiChain scene	Resolved configuration	Solver type
Independent rigid bodies only	`XPBDSolverCfg`	`xpbd`
One or more articulations, without deformables	`MJWarpSolverCfg`	`mujoco_warp`
Cloth or volume soft bodies, without articulations	`VBDSolverCfg`	`vbd`
Articulations together with cloth or volume soft bodies	`DexUniSolverCfg`	`dexuni`

Add every initial robot and object before prepare() so selection sees the complete scene. An explicit {"solver_type": "auto"} or {"class_type": "AutoSolverCfg"} mapping is equivalent to leaving solver_cfg unset. After preparation, inspect sim.physics.solver_type; do not treat physics: newton alone as a solver qualification.

For native solver feature comparisons, joint support, contact-material support, and algorithm details, use the Newton Solver Guide instead of inferring those properties from the table above. The table describes DexSim’s scene routing as exposed by this EmbodiChain integration.

Explicit solver selection#

Choose a concrete solver when an algorithm or solver-specific parameter must be fixed. EmbodiChain mapping configs recognize auto, dexuni, mujoco_warp (or mjwarp), xpbd, semi_implicit, featherstone, and vbd. A DexSim NewtonSolverCfg object can also be assigned directly.

AutoSolver does not select a differentiable solver. requires_grad=True therefore requires an explicit semi_implicit configuration. See Newton Runtime Qualification before treating a successfully decoded configuration as a supported runtime path.

Configuration reference#

Scene parameters belong to NewtonPhysicsCfg; per-object collision, material, and joint properties belong to asset configurations. Default-backend fields such as gpu_memory, bounce_threshold, and enable_ccd do not belong here.

Parameter	Default	Meaning
`physics_dt`	`0.01`	One EmbodiChain physics step in seconds.
`device`	`"cuda:0"`	Compute device; explicit CPU execution remains solver/asset dependent.
`gravity`	`[0.0, 0.0, -9.81]`	World-frame acceleration in m/s².
`num_substeps`	`10`	Solver substeps per physics step; interval is `physics_dt / num_substeps`.
`solver_cfg`	`None`	Preserve DexSim AutoSolver; a mapping or native config fixes a solver.
`collision_cfg`	`NewtonCollisionPipelineCfg()`	External Newton collision pipeline. Set `None` when the selected path owns collision generation, including DexUni.
`requires_grad`	`False`	Enable the current differentiable integration; requires explicit `semi_implicit`.
`use_cuda_graph`	`True`	Request graph capture when supported; unavailable on CPU and disabled in gradient mode.
`sync_to_renderer`	`None`	Preserve DexSim’s automatic render-sync policy; cameras still synchronize on demand.
`debug_mode`	`False`	Enable additional runtime diagnostics.
`suppress_warp_kernel_logs`	`True`	Suppress Warp startup and kernel compilation chatter, but not warnings or errors.

The external collision_cfg only matters when the resolved solver consumes that pipeline. It does not force a solver to use those contacts. Tune collision capacity, refresh intervals, and margins only after confirming the active collision path.

What to read next#

DexSim Unified Coupled Solver For Embodied AI — DexUni capabilities, configuration, and current limitations.
Newton Runtime Qualification — solver/device capability matrix and post-prepare() qualification workflow.
Soft Object and Cloth Object — EmbodiChain deformable configuration and runtime state.

For Newton-native solver internals and tuning, refer directly to the official solver documentation and solver tuning reference.

Newton topics