Grasp pose detection
Grasp candidate generation
(1) Theoretical framework
It is typically based on point contacts that define what components of contact forces and torques (i.e. wrenches) can be transmitted at a specific contact and act on the object. The limitation of these analytical approaches is that they assume full knowledge of object shape and geometry, material properties, and dynamics parameters. But in reality, this information is rarely directly observable but can only be inferred from partial, noisy sensory data.
(2) Deep learning techniques
A. Sampling method
Implementation:
This method involves generating a set of potential grasps, evaluating each one using a deep neural network that estimates the quality of the grasp. The sampling can be done in various spaces such as Euclidean space, latent space, or using priors.
Advantage:
Adjustable processing speed: The number of samples or the level of optimization can be easily modified. Flexibility: Quickly generates a wide array of grasps of varying quality, which increases the chances of finding a high-quality grasp.
Disadvantage:
- Lack of real-time capability: None of the sampling methods have been found to have real-time capabilities.
- Computational cost: A large number of grasps may be needed to find a high-quality grasp, which can be computationally expensive.
Popular method: Dex-net 3.0: Computing robust vacuum suction grasp targets in point clouds using a new analytic model and deep learning
B. Direct regression method
Implementation: Utilize a single network to process the entire input to regress the output (end-to-end method)
Advantage:
Real-time capability:
- Some direct regression methods have demonstrated the ability to run in real-time. End-to-end learning: They can learn the mapping from input to output directly, reducing the need for manual feature engineering.
Disadvantage:
- Dimensionality curse: The difficulty of the problem increases with the dimensionality of the output, such as a 6-DoF grasp pose.
- Over-simplification: May assume there is only one optimal grasp for each input, ignoring the potential for multiple successful grasps.
Popular method: Graspnet-1billion: a large-scale benchmark for general object grasping
C. Reinforcement Learning
Implementation: RL methods learn a policy to maximize a cumulative reward function based on the robot's actions.
Advantage: Policy optimization: Learns the optimal grasping policy through interaction with the environment. Considers trajectory: Focuses not only on grasp generation but also on the trajectory used by the system.
Disadvantages: Data inefficiency: Typically requires a large amount of data to train an effective policy. Single grasp generation: Generates only one grasp per scene, making it difficult to incorporate additional constraints during execution.
D. Exemplar method It rely on a knom m n n n n n nmmnnnnnnwledge database of objects with similar shapes, was the least commonly used approach.
- GraspNet-1Billion: A Large-Scale Benchmark for General Object Grasping
- Graspness Discovery in Clutters for Fast and Accurate Grasp Detection
- 6-DoF Contrastive Grasp Proposal Network
- UniDexGrasp++: Improving Dexterous Grasping Policy Learning via Geometry-aware Curriculum and Iterative Generalist-Specialist Learning
- Learning High-DOF Reaching-and-Grasping via Dynamic Representation of Gripper-Object Interaction
- Differentiable Force Closure https://sites.google.com/view/ral2021-grasp/home
- Context based grasp posture - Object-Centric Grasping Transferability
- An overview of 3D object grasp synthesis algorithms
- Analytical Grasp Generation
- Unions of Balls for Shape Approximation in Robot Grasping
- Can use tools like opengrasp, syngrasp and graspit to generate grasps.