Equivariance
- Also called: equivariant networks, symmetry-aware learning, group equivariance
- Category: Learning & autonomy › Learning from demonstration
- Tags: geometric-deep-learning, symmetry, robot-learning, sample-efficiency, manipulation
A property of a function or neural network whose output transforms predictably when its input is transformed, for example a grasp prediction that rotates when the scene rotates.
Definition
Equivariance describes a function whose output changes in a matching, predictable way when its input is transformed. Formally, a function f is equivariant to a group of transformations if applying a transformation g before the function gives the same result as applying the corresponding transformation after it: f(g·x) = g·f(x). Invariance is the special case where the output does not change at all: f(g·x) = f(x).
Robotics is full of such symmetries. If a scene on a table is rotated, the correct grasp should rotate with it; if an object is moved, the motion needed to pick it up should move too. The relevant transformation groups are usually rotations and translations in the plane (SO(2), SE(2)) or in three dimensions (SO(3), SE(3)). Ordinary convolutional networks are already equivariant to image translations, which is one reason they work well for vision; equivariant networks extend this to rotations and other transformations.
Equivariant architectures build the symmetry into the network's layers rather than hoping the network learns it from data. Group-equivariant convolutional networks, steerable convolutions and point-cloud methods such as vector neurons are examples. Because the network does not need separate examples of every orientation, equivariant models can learn from fewer demonstrations and generalize to poses never seen in training.
The main design question is which symmetry actually holds. Gravity and the table surface break full 3D rotational symmetry, so many manipulation systems use rotation about the vertical axis only. Choosing a symmetry that the task does not have forces the model to produce wrong answers, while ignoring a symmetry that does hold wastes data.
Why it matters
Robot data is expensive to collect, and equivariance is one of the most direct ways to get more generalization from the same amount of data. It lets a policy trained on a few object poses act sensibly on rotated or shifted versions of the same scene.
Key applications
- Grasp and pick-and-place prediction from images or point clouds
- Sample-efficient policy learning from small numbers of demonstrations
- Pose estimation and object-centric perception
- Assembly and placement tasks where part orientation varies
In practice
A team training a tabletop manipulation policy from demonstrations might use a network equivariant to planar rotations, so that a demonstration recorded with a mug facing one way also teaches the policy what to do when the mug faces another. They first check which symmetries the task really has, for example whether the gripper or the camera mounting breaks rotational symmetry, then choose an equivariant library and compare against a standard network trained with rotation data augmentation as a baseline.
Common misconceptions
- Equivariance and invariance are the same — an invariant output stays fixed under a transformation, while an equivariant output transforms along with the input
- Data augmentation gives the same result — augmentation encourages a symmetry approximately, while an equivariant architecture enforces it exactly
- More symmetry is always better — enforcing a symmetry the task does not have, such as full 3D rotation under gravity, forces the model to be wrong
Who should know this
- Robot learning researchers working on manipulation policies
- Perception engineers working with point clouds and pose estimation
- Machine learning engineers designing network architectures for robotics
Related technologies
- Group-Equivariant CNNs
- Steerable CNNs
- Vector Neurons
- Data Augmentation
- Diffusion Policy
- Lie Groups (SE(3))
Related terms
Related projects
- LeRobot: HuggingFace ML library for real-world robotics in PyTorch
- OpenPI: Physical Intelligence VLA foundation models for general robot control
- Isaac Lab: NVIDIA robot learning framework: one task API across Newton, PhysX and Isaac Sim
- NVIDIA Isaac GR00T: NVIDIA open foundation model for humanoid robot reasoning and skills
- OpenVLA: 7B-parameter Vision-Language-Action model for robotic manipulation
- Drake: MIT robotics toolbox for model-based dynamics, optimization, and control system design
Videos
- Equivariant Neural Networks | Part 1/3 - Introduction: DeepFindr