Principle:Haosulab ManiSkill Rotation Geometry Operations
| Knowledge Sources | |
|---|---|
| Domains | Robotics, Simulation, Computational_Geometry |
| Last Updated | 2026-02-15 08:00 GMT |
Overview
Spatial math encompasses the rotation representations, coordinate transformations, and geometric computations required to reason about object positions, orientations, and spatial relationships in three-dimensional robotics simulation.
Description
Every aspect of robotics simulation involves spatial reasoning: where is the robot's end-effector relative to the target object? What orientation must a grasped object be rotated to? How far has a valve been turned? These questions are answered by a library of spatial math operations that convert between rotation representations, compute bounding volumes, extract geometric properties from mesh data, and perform spatial transformations on batched tensors.
The Spatial Math principle provides four categories of functionality. First, rotation conversions handle the many ways rotations are represented -- quaternions (wxyz), rotation matrices (3x3), axis-angle vectors, Euler angles (multiple conventions) -- and convert freely between them. This is essential because different components of the system prefer different representations: the physics engine uses quaternions, neural networks often use rotation matrices or 6D representations, and human interfaces use Euler angles. Second, geometry utilities compute spatial relationships such as distances between points and lines, bounding box dimensions, transformations between coordinate frames, and point cloud operations. Third, bounding cylinder computation fits minimal enclosing cylinders to point sets, which is used for object shape approximation in reward computation. Fourth, mesh utilities extract collision geometry from simulation objects and convert it to standard mesh formats for analysis and visualization.
All operations are implemented as differentiable tensor operations (using PyTorch or NumPy), supporting both CPU and GPU execution and enabling gradient flow through spatial computations when needed for differentiable physics or learned controllers.
Usage
This principle applies whenever:
- Rotation representations must be converted (e.g., quaternion to rotation matrix for a neural network input, or Euler angles to quaternion for a reward function).
- Spatial relationships must be computed in batched form across thousands of parallel environments (distances, relative poses, angle differences).
- Object geometry must be approximated by bounding volumes (cylinders, boxes) for reward shaping or collision checking.
- Collision meshes must be extracted from simulation objects for visualization, analysis, or motion planning.
Theoretical Basis
1. Rotation Representations:
- Quaternion (wxyz): A 4D unit vector representing a rotation. Compact, avoids gimbal lock, efficient for composition. The simulation engine's native representation.
- Rotation Matrix (3x3): A 3x3 orthogonal matrix with determinant +1. Natural for transforming vectors and composing rotations via matrix multiplication. Used in neural network architectures.
- Axis-Angle: A 3D vector whose direction is the rotation axis and magnitude is the rotation angle. Intuitive for small rotations and angular velocity.
- Euler Angles: Three angles representing sequential rotations about coordinate axes (e.g., XYZ, ZYX). Intuitive for human specification but subject to gimbal lock.
- 6D Representation: The first two columns of the rotation matrix, from which the full matrix can be recovered via Gram-Schmidt orthogonalization. Continuous and free of singularities, making it well-suited for neural network output.
2. Conversion Graph: The conversion functions form a directed graph where any representation can be converted to any other. Key conversions include quaternion-to-matrix, matrix-to-quaternion, axis-angle-to-quaternion, and Euler-to-matrix. Each conversion is implemented as a differentiable tensor operation.
3. Bounding Volume Computation: Given a set of 3D points, compute the minimal enclosing cylinder aligned with a principal axis. The algorithm projects points onto a 2D plane, finds the minimal enclosing circle (center and radius), and uses the min/max along the axis for height. This provides a simple shape approximation for elongated objects like pegs and rods.
4. Mesh Extraction: Collision shapes attached to simulation actors (boxes, spheres, capsules, convex meshes, triangle meshes) are converted to a uniform triangle mesh representation. This involves generating vertices and faces for primitive shapes and extracting raw mesh data for complex shapes. The resulting meshes can be used with libraries like Trimesh for analysis.
5. Spatial Transformation: Points and vectors are transformed between coordinate frames using rigid body transformations (rotation + translation). The framework supports batched transformation of point clouds, enabling efficient coordinate conversion across all parallel environments.
Related Pages
- Implementation:Haosulab_ManiSkill_RotationConversions -- Comprehensive rotation representation conversion library.
- Implementation:Haosulab_ManiSkill_GeometryUtils -- Spatial geometry utility functions (distances, transforms, bounding boxes).
- Implementation:Haosulab_ManiSkill_BoundingCylinder -- Axis-aligned bounding cylinder computation algorithm.
- Implementation:Haosulab_ManiSkill_TrimeshUtils -- Collision shape to triangle mesh extraction utilities.