OVPhysX FrameTransformer Sensor — Design Spec

Issue: #5320 — [OVPHYSX] FrameTransformer sensor
Date: 2026-04-20
Status: Draft

Summary

Implement FrameTransformer and FrameTransformerData for the OVPhysX backend, satisfying the BaseFrameTransformer and BaseFrameTransformerData contracts. Computes relative transforms between a source frame and multiple target frames attached to rigid bodies. Only requires body transforms (no velocities or accelerations).

Validation environment: Isaac-Lift-Cube-Franka-v0 (manager-based) — uses FrameTransformer to track end-effector pose relative to robot base. Also validates #5322 (manipulation support).

Contract to Satisfy

BaseFrameTransformer

Inherits SensorBase. Additional abstract:

BaseFrameTransformerData

Where N=num_envs, M=num_target_frames.

Architecture

File Layout

source/isaaclab_ovphysx/isaaclab_ovphysx/sensors/
└── frame_transformer/
    ├── __init__.py
    ├── __init__.pyi
    ├── frame_transformer.py       (~500 lines)
    ├── frame_transformer_data.py  (~150 lines)
    └── kernels.py                 (~100 lines)

Implementation Pattern

frame_transformer.py:

1. _initialize_impl() (most complex part):

   • Resolve source prim path and verify rigid body exists
   • Resolve all target frame prim paths and verify rigid bodies
   • Compute fixed offsets for source and target frames (once)
   • Build body tracking: create a combined view of all tracked rigid bodies
   • Build per-env index arrays:
     • _source_raw_indices: maps env → position in raw transform array
     • _target_raw_indices: maps (env, frame) → position in raw transform array
   • Pre-compute offset arrays as GPU warp arrays
   • Allocate output buffers

2. _update_buffers_impl(env_mask):

   • Read all tracked body transforms from ovphysx bindings (single batched read)
   • Launch frame_transformer_update_kernel with 2D grid (num_envs, num_target_frames):
     • Fetch source transform by index, apply source offset
     • Fetch target transform by index, apply target offset
     • Compute relative: target_relative = source_inverse * target_offset
     • Store 6 outputs (source and target poses in world + relative)

OVPhysX API Requirements

No new ovphysx APIs needed — only body transforms are required, which are already available through existing Articulation and (future) RigidObject bindings.

The key implementation question is how to create a single batched view spanning multiple distinct body paths. In PhysX, RigidBodyView accepts a glob expression. For ovphysx, the approach depends on whether TensorBindings can be queried by body index within an articulation, or whether a separate mechanism is needed.

Question for @marcodiiga: How to read transforms for arbitrary bodies (not just the root) from ovphysx? The Articulation's ARTICULATION_BODY_POSE binding returns all link poses — FrameTransformer needs to index into specific links by name.

Warp Kernels

Tests

Source: source/isaaclab_physx/test/sensors/test_frame_transformer.py
Target: source/isaaclab_ovphysx/test/sensors/test_frame_transformer.py

Validation

Isaac-Lift-Cube-Franka-v0:

• FrameTransformer tracks Franka end-effector relative to base
• Verify source/target poses match expected robot kinematics
• Verify relative transforms are consistent with forward kinematics

Dependencies

• Requires PR ovphysx backend integration #4852 merged
• No new ovphysx APIs needed (uses body transforms from Articulation bindings)
• Independent of other sensor specs — can be implemented in parallel

Estimated Scope

• frame_transformer.py: ~500 lines (initialization is complex, update is simple)
• frame_transformer_data.py: ~150 lines
• kernels.py: ~100 lines
• Tests: ~200 lines of adaptations