Glossary
Zero-Shot Manipulation
Robotic manipulation of objects the system has never seen during training — generalizing motor skills to novel geometries, materials, and configurations without additional learning.
Zero-shot manipulation refers to a robotic system's ability to successfully manipulate objects it has never encountered during training. "Zero-shot" means no additional learning steps, fine-tuning, or demonstrations are required when encountering new objects — the system generalizes from its training distribution to genuinely novel instances.
This is one of the hardest open problems in robotics. Physical manipulation depends on contact geometry, friction coefficients, mass distribution, and compliance properties that vary enormously across objects and are difficult to predict without direct physical experience.
Why zero-shot generalization is hard
Most manipulation systems that work well on their training distribution fail on novel objects because they are learning to memorize object-specific solutions rather than learning the underlying physics of manipulation. A policy trained on a specific mug learns representations tied to that mug's geometry — representations that do not transfer to a differently shaped cup, even if the manipulation task is identical.
True zero-shot generalization requires policies that operate in terms of physical primitives — contact forces, friction cones, grasp stability conditions — rather than object-specific features.
How Webbeon approaches Zero-Shot Manipulation
Object Class achieves zero-shot manipulation through several key design choices:
Demonstration-free training: Object Class is trained entirely through reinforcement learning in physics simulation, without human demonstrations. This prevents the policy from overfitting to the specific motion patterns a human demonstrator would use.
Aggressive domain randomization: Object geometry, mass, friction, and compliance are randomized throughout training, forcing the policy to develop representations that capture physically meaningful properties rather than surface-level features.
Tactile feedback integration: 192 taxels per fingertip, combined with vibration sensing for slip detection, give Object Class physical feedback that makes manipulation robust to visual uncertainty and object property variation.
Physics-grounded reward: Rewards tied to physical task completion rather than trajectory matching allow the policy to discover diverse solutions to the same manipulation problem.
Key facts
- 89% first-attempt grasp success rate on objects entirely outside the training distribution
- 147-object benchmark covering household items from tools to irregular shapes
- Object Class discovers manipulation strategies not found in human grasp taxonomies — including finger-gaiting and palm-bracing behaviors that emerge from physics optimization
- Strongest imitation-learning baseline achieves 61% on the same benchmark