Batched simulations

Some applications need to simulate many independent scenes at the same time. This is typically used for evaluating several robot controllers or running many copies of a training environment. A NexusState groups these rigid-body scenes into a batch, so NexusPipeline can process them together.

Creating environments

A new NexusState already contains environment 0. Each call to state.add_environment creates one additional environment and returns its index. Call state.insert_rigid_body_in with this index:

Example 2D

let mut state = NexusState::default();
state.set_rbd_collisions_capacity(128);
for i in 0..4 {
    let env = if i == 0 { 0 } else { state.add_environment() };
    let body = RigidBodyBuilder::dynamic()
        .translation(Vec2::new(i as f32 * 2.0, 3.0))
        .build();
    let collider = ColliderBuilder::ball(0.5).build();
    state.insert_rigid_body_in(env, body, collider, RbdCoupling::None);
}

Example 3D

let mut state = NexusState::default();
state.set_rbd_collisions_capacity(128);
for i in 0..4 {
    let env = if i == 0 { 0 } else { state.add_environment() };
    let body = RigidBodyBuilder::dynamic()
        .translation(Vec3::new(i as f32 * 2.0, 3.0, 0.0))
        .build();
    let collider = ColliderBuilder::ball(0.5).build();
    state.insert_rigid_body_in(env, body, collider, RbdCoupling::None);
}

state = NexusState()
state.set_rbd_collisions_capacity(128)
for i in range(4):
    env = 0 if i == 0 else state.add_environment()
    body = RigidBodyBuilder.dynamic().translation(Vec3(i * 2.0, 3.0, 0.0)).build()
    collider = ColliderBuilder.ball(0.5).build()
    state.insert_rigid_body_in(env, body, collider, RbdCoupling.NONE)

Objects in different environments never collide. The RigidBodyBuilder::translationRigidBodyBuilder.translation values above only separate the balls for visualization; identical positions in separate environments would work too.

The RigidBodyHandle returned by state.insert_rigid_body_in belongs to that environment. Keep the environment index with each handle, and pass the same index to state.insert_collider_in, state.insert_impulse_joint_in, or state.insert_multibody_joint_in when referring to it.

Matching topology

All environments passed to state.finalize must have matching structure:

• The same counts of RigidBody and Collider values.
• The same counts of impulse joints and multibodies.
• Compatible joint topology, constructed in the same order.
• Identical RbdSimParams values, including num_solver_iterations.

Initial poses, velocities, and shapes may differ. It is recommended to call the same scene-building function for every environment, changing only the initial conditions that you want to vary. Populate every environment before calling state.finalize, so an intermediate environment doesn't remain empty.

warning

Calling state.add_environment without populating the new environment like the others makes state.finalize fail. The batch layout doesn't automatically add missing objects. Check each environment when finalization reports incompatible counts or parameters.

Batching applies to the rigid-body solver. Calling state.add_environment doesn't duplicate the MPM state; see MPM coupling and environments for that limitation.

Rendering a batch

Use viewer.insert_shape_in(env, handle, &shape, local_pose)viewer.insert_shape_in(env, handle, shape, local_pose) to register graphics for a body in a particular environment. Calling viewer.insert_shape instead targets environment 0. The local_pose uses the same frame as the collider's local position.

The same viewer.sync call synchronizes all environments. Rendering a large batch can still take time even if the physics scenes are small, so measure pipeline.simulate separately from viewer.sync and viewer.render_frame when evaluating simulation throughput.

Memory usage

Call state.set_rbd_collisions_capacity before state.finalize to set the collision-pair capacity per environment. The example uses 128 because each environment contains only one body. Workspaces scale with the environment count: a capacity appropriate for one large scene may waste memory for thousands of small scenes.

A scene with many contacts needs a larger state.set_rbd_collisions_capacity value. NexusPipeline can grow collision buffers after overflow, but it cannot restore contacts omitted during that step. NexusCapacities::rbd_resize_policy configures this recovery policy when constructing NexusState. Choose sufficient capacity for the expected workload before increasing the number of environments.

Resetting episodes

For repeated episode resets in 3D, RbdState::snapshot and RbdState::reset_env_from_snapshot restore a matching environment without rebuilding the whole scene. For large batches, RbdState::publish_reset_templates and RbdState::reset_envs_from_templates keep reset templates on the GPU. See the reset example. The current Python bindings don't expose RbdSnapshot or GPU reset templates. Construct and populate a fresh NexusState, then call state.finalize to restart the batch from its authored scene.

Ordinary state.insert_rigid_body_in calls change the authored scene. Calling them after an episode has started can restore all bodies to their initial positions on the next state.finalize. See changing the scene description before modifying a running batch.