[OVPHYSX] RigidObjectCollection asset

[AntoineRichard]

Summary

Implement RigidObjectCollection and RigidObjectCollectionData for the OVPhysX backend. Extends RigidObject to manage multiple distinct rigid bodies per environment with dual (env, body) indexing.

Validation environment: Franka cube stacking task — multiple cubes (RigidObjectCollection) + Franka arm (Articulation).

Scope

• rigid_object_collection.py (~1200 lines), rigid_object_collection_data.py (~800 lines)
• View-order reshaping kernels (body-major ↔ instance-major)
• Copy PhysX tests + add OVPhysX to interface tests

Key Complexity

Dual indexing with view-order reshaping: methods accept (env_ids, body_ids) but ovphysx bindings use flat body-major ordering. Reshape kernels handle the mapping.

OVPhysX API Requirements

Confirm ovphysx TensorBindings can create a single view spanning multiple distinct rigid body prims (not just clones). @marcodiiga to verify.

Dependencies

• Requires [OVPHYSX] RigidObject asset #5316 (RigidObject) — shares kernels and patterns
• Requires PR ovphysx backend integration #4852 merged

Spec

Full design spec: `docs/superpowers/specs/2026-04-20-ovphysx-rigid-object-collection-design.md`

Parent issue: #5315

---

📋 Full Design Spec (click to expand)

OVPhysX RigidObjectCollection — Design Spec

Issue: #5317 — [OVPHYSX] RigidObjectCollection asset
Date: 2026-04-20
Status: Draft

Summary

Implement RigidObjectCollection and RigidObjectCollectionData for the OVPhysX backend, satisfying the BaseRigidObjectCollection and BaseRigidObjectCollectionData contracts. This extends the single-body RigidObject to manage multiple distinct rigid bodies per environment, adding dual (env, body) indexing.

Validation environment: Franka cube stacking task (stack_joint_pos_instance_randomize_env_cfg.py) — uses a Franka arm (Articulation) and multiple cubes (RigidObjectCollection).

Depends on: #5316 (RigidObject) — shares kernels and patterns.

Guiding Principles

Same as RigidObject spec:

• Mirror PhysX structure
• Reuse patterns from OVPhysX Articulation and RigidObject
• Do not modify contracts
• Tests copied with only setup changes

Key Difference from RigidObject

RigidObjectCollection manages N distinct rigid bodies across M environments. This adds:

1. Dual indexing — methods accept both env_ids and body_ids (or env_mask and body_mask)
2. View-order reshaping — ovphysx tensor bindings likely use body-major ordering [body0_env0, body0_env1, ..., body1_env0, ...]. Methods must reshape between instance-order (num_instances, num_bodies, ...) and view-order (num_bodies * num_instances, ...)
3. Per-body properties — mass, inertia, COM are independent per body per environment
4. Data dimensionality — all state tensors are (num_instances, num_bodies, ...) instead of (num_instances, ...)

Contract to Satisfy

BaseRigidObjectCollection

Abstract properties (7): Same as RigidObject — data, num_instances, num_bodies, body_names, root_view, instantaneous_wrench_composer, permanent_wrench_composer.

Abstract methods — core (4):

• reset(env_ids, object_ids, env_mask) — note extra object_ids parameter vs RigidObject
• write_data_to_sim()
• update(dt)
• find_bodies(name_keys, preserve_order) -> (torch.Tensor, list[str]) — returns Tensor not list[int]

Abstract methods — body pose writers (6):

• write_body_pose_to_sim_index/mask(body_poses, body_ids, env_ids)
• write_body_link_pose_to_sim_index/mask(body_poses, body_ids, env_ids)
• write_body_com_pose_to_sim_index/mask(body_poses, body_ids, env_ids)

Abstract methods — body velocity writers (6):

• write_body_velocity_to_sim_index/mask(body_velocities, body_ids, env_ids)
• write_body_link_velocity_to_sim_index/mask(body_velocities, body_ids, env_ids)
• write_body_com_velocity_to_sim_index/mask(body_velocities, body_ids, env_ids)

Abstract methods — body property setters (6):

• set_masses_index/mask(masses, body_ids, env_ids)
• set_coms_index/mask(coms, body_ids, env_ids)
• set_inertias_index/mask(inertias, body_ids, env_ids)

Abstract methods — deprecated (3):

• write_body_state_to_sim, write_body_com_state_to_sim, write_body_link_state_to_sim

Internal hooks (4):

• _initialize_impl, _create_buffers, _process_cfg, _invalidate_initialize_callback

BaseRigidObjectCollectionData

Abstract properties (~30): Body state (link/com pose, velocity, acceleration), body properties (mass, inertia, com), derived quantities (projected_gravity, heading, body-frame velocities), sliced components. All shapes are (num_instances, num_bodies, ...).

Abstract method (1): update(dt: float)

Architecture

File Layout

source/isaaclab_ovphysx/isaaclab_ovphysx/assets/
├── __init__.py          (add RigidObjectCollection exports)
├── __init__.pyi
├── articulation/        (existing)
├── rigid_object/        (from #5316)
└── rigid_object_collection/
    ├── __init__.py
    ├── __init__.pyi
    ├── rigid_object_collection.py       (~1200 lines)
    └── rigid_object_collection_data.py  (~800 lines)

Implementation Pattern

rigid_object_collection.py:

1. _initialize_impl():

   • Get OvPhysxManager instance
   • Resolve all rigid body prim paths (multiple bodies)
   • Create TensorBindings for all bodies as a single batched view
   • Build _env_body_ids_to_view_ids() mapping for dual indexing
   • Create RigidObjectCollectionData with lazy binding getter

2. View-order reshaping (critical complexity):

   • reshape_view_to_data(view_tensor) — convert (num_bodies * num_instances, D) to (num_instances, num_bodies, D)
   • reshape_data_to_view(data_tensor) — reverse
   • _env_body_ids_to_view_ids(env_ids, body_ids) — map (env_id, body_id) pairs to flat view indices
   • These reshaping functions use Warp kernels for GPU-native operation

3. Body state writers — same pattern as RigidObject but with dual indexing:

   • Accept (num_selected_envs, num_selected_bodies, D) input
   • Map to view-order flat indices
   • Write via TensorBinding with flat indices

4. Wrench application — wrenches applied per (env, body) pair, reshaped to view order

rigid_object_collection_data.py:

• Same lazy binding + timestamped caching pattern
• All properties return (num_instances, num_bodies, ...) tensors
• Derived properties (gravity projection, heading, body-frame velocities) computed per body

OVPhysX API Requirements

Same tensor types as RigidObject, but the bindings must handle multiple bodies:

• The ovphysx binding shape becomes (num_bodies * num_instances, D) in view order
• Need to confirm ovphysx supports batched rigid body views across multiple distinct prims

Blocker for @marcodiiga: Confirm that ovphysx TensorBindings can create a single view spanning multiple distinct rigid body prims (not just clones of the same prim). This is how PhysX's RigidBodyView works with a glob expression matching multiple body paths.

Warp Kernels

In addition to kernels shared with RigidObject:

Kernel	Purpose
reshape_view_to_data_2d	Reorder body-major to instance-major
reshape_data_to_view_2d	Reorder instance-major to body-major
env_body_to_view_scatter	Scatter write for dual-indexed updates

Tests

Backend-specific tests (copy from PhysX)

Source: source/isaaclab_physx/test/assets/test_rigid_object_collection.py
Target: source/isaaclab_ovphysx/test/assets/test_rigid_object_collection.py

Changes: swap PhysxCfg for OvPhysxCfg, adapt USD asset paths.

Interface tests

File: source/isaaclab/test/assets/test_rigid_object_collection_iface.py

Add "ovphysx" backend parametrization with mock bindings.

Validation

Franka cube stacking — exercises RigidObjectCollection with multiple cubes, Franka arm (Articulation), and FrameTransformer sensor.

Validation steps:

1. Run stacking env with OVPhysX for 100 steps
2. Verify each cube in the collection has independent state
3. Verify dual-indexed resets (reset specific cubes in specific envs)
4. Verify body property reads return correct per-body values

Dependencies

• Requires [OVPHYSX] RigidObject asset #5316 (RigidObject) — shares kernels and patterns
• Requires PR ovphysx backend integration #4852 merged
• May require ovphysx wheel update for multi-body rigid body views

Estimated Scope

• rigid_object_collection.py: ~1200 lines
• rigid_object_collection_data.py: ~800 lines
• Reshape kernels: ~100 lines
• Tests: ~250 lines of adaptations