search

perezjln / gym-lowcostrobot

Other
29 stars 18 forks source link

Todos #2

Closed qgallouedec closed 5 months ago

qgallouedec commented 6 months ago
perezjln commented 6 months ago

I take what should be the simplest, support for rendering first ... and I should finally learning to use got properly ^^"

https://github.com/perezjln/gym-lowcostrobot/commit/416c8e67bc50f3104372e73e202f21404cb29c14 https://github.com/perezjln/gym-lowcostrobot/commit/c3ab416896274f4218ee342100b177976fab32cd

perezjln commented 6 months ago

Find the correct fps : it seems it can be set, not sure need to recheck on documentation. https://github.com/perezjln/gym-lowcostrobot/commit/c3ab416896274f4218ee342100b177976fab32cd

perezjln commented 6 months ago

Regarding determinism of the environment, maybe @qgallouedec you have an example to share ?

One point I'm wondering is how to retrieve poses from mujoco, it seems two ways are somehow possible:

https://github.com/perezjln/gym-lowcostrobot/blob/c3ab416896274f4218ee342100b177976fab32cd/gym_lowcostrobot/envs/lift_cube_env.py#L61

or

https://github.com/perezjln/gym-lowcostrobot/blob/c3ab416896274f4218ee342100b177976fab32cd/examples/mujoco_simple_invk.py#L35

perezjln commented 6 months ago

Add and test actuator ranges retrieval in base env. https://github.com/perezjln/gym-lowcostrobot/commit/fbcafd90f06745b4122ec67fa2904745cd08c619

Not sure how to estimate ranges for the object part of the observation space, one idea @qgallouedec ?

qgallouedec commented 6 months ago

Find the correct fps : it seems it can be set, not sure need to recheck on documentation. c3ab416

For the moment, I'd opt for a fixed fps, which the user can't set. It has to be part of the environment's metadata.

Regarding determinism of the environment, maybe @qgallouedec you have an example to share ?

import gymnasium as gym
import gym_lowcostrobot

env = gym.make('LiftCube-v0')

observation1, info = env.reset(seed=123)
observation2, info = env.reset(seed=123)

assert (observation1 == observation2).all()

that why the CI fails.

One point I'm wondering is how to retrieve poses from mujoco, it seems two ways are somehow possible:

https://github.com/perezjln/gym-lowcostrobot/blob/c3ab416896274f4218ee342100b177976fab32cd/gym_lowcostrobot/envs/lift_cube_env.py#L61

or

https://github.com/perezjln/gym-lowcostrobot/blob/c3ab416896274f4218ee342100b177976fab32cd/examples/mujoco_simple_invk.py#L35

We should draw inspiration from gym_aloha, I'm not familiar with MuJoCo.

qgallouedec commented 6 months ago

Add and test actuator ranges retrieval in base env. fbcafd9

The goal of having this range is to specify the action_space right?

Not sure how to estimate ranges for the object part of the observation space, one idea @qgallouedec ?

It doesn't have to be precise, but it would have to be at least one of the finite bounds for the observation space.

perezjln commented 6 months ago

Add and test actuator ranges retrieval in base env. fbcafd9

The goal of having this range is to specify the action_space right?

The goal of defining the ranges is rather for the observation space actually. The observation is composed with the joints pose/velocity/accel associated with each actuator + same info for each object in the scene. Having the range will allow to normalize as we know.

The action space are rather displacement than absolute pose in both cases (joint and Cartesian), so they can be defined a priori.

Not sure how to estimate ranges for the object part of the observation space, one idea @qgallouedec ?

It doesn't have to be precise, but it would have to be at least one of the finite bounds for the observation space.

Not obvious to me in the case of the objects to manipulate , maybe I should check how you did in panda-gym.

perezjln commented 6 months ago

Regarding determinism of the environment, maybe @qgallouedec you have an example to share ?

import gymnasium as gym
import gym_lowcostrobot

env = gym.make('LiftCube-v0')

observation1, info = env.reset(seed=123)
observation2, info = env.reset(seed=123)

assert (observation1 == observation2).all()

that why the CI fails.

Thanks, I will check this…. Interesting

One point I'm wondering is how to retrieve poses from mujoco, it seems two ways are somehow possible: https://github.com/perezjln/gym-lowcostrobot/blob/c3ab416896274f4218ee342100b177976fab32cd/gym_lowcostrobot/envs/lift_cube_env.py#L61

or https://github.com/perezjln/gym-lowcostrobot/blob/c3ab416896274f4218ee342100b177976fab32cd/examples/mujoco_simple_invk.py#L35

We should draw inspiration from gym_aloha, I'm not familiar with MuJoCo.

Good idea, I just check, they use the qpos from the mujoco.data: https://github.com/huggingface/gym-aloha/blob/7f19312f453f45fbe8ca8883d6250626dfa36373/gym_aloha/tasks/sim.py#L61

It makes sense as it seems to work in the invk script of our examples module.

qgallouedec commented 6 months ago

The goal of defining the ranges is rather for the observation space actually. The observation is composed with the joints pose/velocity/accel associated with each actuator + same info for each object in the scene. Having the range will allow to normalize as we know.

Makes sense!

Not obvious to me in the case of the objects to manipulate , maybe I should check how you did in panda-gym.

I will check, I can't remember tbh.

perezjln commented 6 months ago

The goal of defining the ranges is rather for the observation space actually. The observation is composed with the joints pose/velocity/accel associated with each actuator + same info for each object in the scene. Having the range will allow to normalize as we know.

Makes sense!

Not obvious to me in the case of the objects to manipulate , maybe I should check how you did in panda-gym.

I will check, I can't remember tbh.

It seems like the position ranges for the object is pre-defined: https://github.com/qgallouedec/panda-gym/blob/51fb901db6968cb8a78ebb11070d313566700d38/panda_gym/envs/tasks/push.py#L24

I just implemented it the same way: https://github.com/perezjln/gym-lowcostrobot/commit/8e29cbb470c1e8d0f7d9cc1da304236dc381501b