Improving Grasp Classification through Spatial Metrics Available from Sensors
Improving Grasp Classification through Spatial Metrics Available from Sensors
Embedded Files
We evaluated the usefulness of different sensors for grasp classification using an ablation study and evaluated their robustness to noise and the transfer to the real world.
We evaluated the usefulness of different sensors for grasp classification using an ablation study and evaluated their robustness to noise and the transfer to the real world.
We built a full dataset of 135,000 grasps of 8 different shapes of 3 different sizes and 3 different orientations in Mujoco to train grasp classifiers with. We then built a series of grasp classifiers based on a collection of 76 metrics to evaluate which sensors are most useful for grasp classification. After finding that the majority of the information can be contained within 3 metrics, we trained the classifier with different sensor noise and applied those trained classifiers in the real world to see the effect of noise training on transfer across the sim to real gap. This was presented at ICRA 2021. https://ieeexplore.ieee.org/document/9561750
We built a full dataset of 135,000 grasps of 8 different shapes of 3 different sizes and 3 different orientations in Mujoco to train grasp classifiers with. We then built a series of grasp classifiers based on a collection of 76 metrics to evaluate which sensors are most useful for grasp classification. After finding that the majority of the information can be contained within 3 metrics, we trained the classifier with different sensor noise and applied those trained classifiers in the real world to see the effect of noise training on transfer across the sim to real gap. This was presented at ICRA 2021. https://ieeexplore.ieee.org/document/9561750
Real world trials used a physical Kinova Jaco 2 hand. We used Aruco markers to determine the state of the hand and object that were affixed with custom removable 3d printed mounts.
I built the Mujoco simulation and implemented all the metrics in both the simulator and the real world. I also mentored two undergraduate students who trained random forest classifiers used in the paper and handled the noise generation.
I built the Mujoco simulation and implemented all the metrics in both the simulator and the real world. I also mentored two undergraduate students who trained random forest classifiers used in the paper and handled the noise generation.
Page updated
Google Sites
Report abuse