crowdAI is shutting down - please read our blog post for more information
Reinforcement learning with musculoskeletal models
Stanford Neuromuscular Biomechanics Laboratory
EPFL Digital Epidemiology Lab
over 1 year ago
Looking at the ProstheticsEnv, it looks like only the X and Y position of the body parts are returned in the observation vector. Was this intended? Specifically, these lines are responsible:
cur = 
cur += state_desc["body_pos"][body_part][0:2]
cur += state_desc["body_vel"][body_part][0:2]
cur += state_desc["body_acc"][body_part][0:2]
The full 3D state of the system can still be inferred from the joint angles, but it seems like we should also have Z for the body part coordinates?
over 1 year ago |
You have the freedom to arrange your own observation vector from the dictionary describing the state. Our implementation is just an example.
In particular, the environment in the grader returns dictionaries as observations and not vectors.
We will make it more explicit in the description of the challenge and probably remove the vector observation completely (providing some extra ‘helper’ functions)
Having said that, we didn’t intend to skip Z, thank you for catching this!