GraspFlow mainly tackles single item in the environment.
- E_TYPE: single
- M_TYPE: GraspFlow, graspnet, metropolis
- (cat, idx):
- (box,14)
- (box,17)
- (mug,2)
- (mug,8)
- (mug,14)
- (bottle,3)
- (bottle,12)
- (bottle,19)
- (bowl,1)
- (bowl,16)
- (cylinder,2)
- (cylinder,11)
- (fork, 1)
- (fork, 11)
- (hammer, 15)
- (pan, 3)
- (pan, 6)
- (scissor, 4)
- (scissor, 7)
- (spatula, 1)
- (spatula, 14)
GoES tackles objects in complex environments.
-
E_TYPE: shelf008
-
(cat, idx):
- (bottle 14)
- (bowl 8)
- (bowl 10)
- (pan 6)
- (pan 12)
- (fork 6)
- (scissor 7)
-
E_TYPE: diner001
-
(cat, idx):
- (pan, 12)
- (spatula, 14)
- (bottle, 0)
- (bowl, 8)
- (fork, 6)
For both environments, GoES can optimized using following parameters:
- Classifiers:
- S - Stability Classifier - assesses stability of the optimized grasps.
- E - Executable Classifier - assesses wether grasp lies within robot's reachable map and avoids singularity.
- C - Collision Classifier - assess collision between the grasp and environment.
- N - Intent Classifier - assesses intent affordance for the query. Note: classifier parameter in GoES can be build using any combination of the classifiers above. E.g: SE, SC, SEC, ...
Use config file to indicate formula ranking and other parameters of the GoES.
- optimizer - pytorch optimizer. Keep it as SGD.
- eta_t - learning rate for translation.
- eta_r - learning rate for orientation.
- grad_normalize - boolean indicator responsible for normalization of gradients.
- GoES
- num_samples_per_grasp: number of additional samples per grasp
- grad_iterations: number of lower bound local optimizations
- t_std_dev: standard deviation for translations for ES part of GoES.
- e_std_dev: standard deviation for orientations for ES part of GoES.
- S_warmup_iterations: number of initial iterations for S classifier as a warmup.