Robot Hand Grasp Learning With Symmetry-Aware Posture Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning for grasping postures of robot hands with symmetry can destabilize estimation accuracy due to rotational symmetry, leading to incorrect outputs at symmetry boundaries.
Innovation Solution
A learning apparatus that utilizes machine learning with training data representing a robot hand's 2-fold rotational symmetry, combining a first posture and its 180° rotation, expressed as a single parameter set, to improve estimation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning is performed using training data representing postures of a robot hand with 2-fold rotational symmetry, then the robot hand can utilize its symmetric structure for grasping, but the symmetry destabilizes machine learning and reduces estimation accuracy at symmetry boundaries
Solution Approach 1:
The patent applies asymmetry by introducing a phase shift parameter that breaks the rotational symmetry in the parameter space. By representing the first posture with parameters (θ, φ) and the second posture with parameters (θ + π, φ + δ) where δ is a non-zero phase shift, the training data avoids symmetric boundary conditions that cause instability. This asymmetric parameterization allows the neural network to learn distinct representations for symmetric postures, preventing the destabilization effect while maintaining the robot hand's physical symmetry and grasping versatility.
2Adaptability or versatility
If training data includes both first posture and second posture (180° rotation) as separate entries, then comprehensive grasping scenarios are covered, but incorrect grasping posture may be output at the interface between the two regions
Solution Approach 1:
The patent resolves this contradiction by applying asymmetric parameterization to the training data. Instead of treating the first posture (θ, φ) and second posture (θ + π, φ) as symmetric equivalents that cause boundary confusion, the system represents them with asymmetric parameters (θ, φ) and (θ + π, φ + δ) where δ ≠ 0. This ensures that each posture has a unique parameter representation, eliminating the interface problem between symmetric regions while maintaining comprehensive coverage of all grasping scenarios.
3Ease of manufacture
If the robot hand has 2-fold rotational symmetry, then the structure is simplified and easier to manufacture, but the symmetry causes machine learning instability
Solution Approach 1:
The patent maintains the physical symmetry of the robot hand structure for ease of manufacture while introducing asymmetry in the parameter space for machine learning. The physical robot hand retains its 2-fold rotational symmetry, but the training data uses asymmetric parameterization where the second posture is represented as (θ + π, φ + δ) instead of (θ + π, φ). This decoupling allows the simplified symmetric structure to be manufactured while the asymmetric parameter representation prevents machine learning instability.
Data Source
AI summary
To realize a learning apparatus which improves the estimation accuracy of a grasping posture of a robot hand having symmetry. A learning apparatus according to one embodiment of the present disclosure includes a learning unit configured to learn, by machine learning, a posture for grasping an object by a robot hand, the machine learning being performed using training data represented by one parameter set which includes a first posture of the robot hand having a 2-fold rotational symmetry property and a second posture of the robot hand rotated 180° around an axis of rotational symmetry of the first posture.