Farama-Foundation / Gymnasium

An API standard for single-agent reinforcement learning environments, with popular reference environments and related utilities (formerly Gym)
https://gymnasium.farama.org
MIT License
7.31k stars 817 forks source link

[Bug Report] No reward threshold for Walker2d-v4 #1157

Closed modanesh closed 2 months ago

modanesh commented 2 months ago

Describe the bug

All MuJoCo envs have reward threshold except for Walker2d-v4:

>>> import gymnasium as gym
>>> gym.__version__
'0.29.1'
>>> gym.make("Ant-v4").spec.reward_threshold
6000.0
>>> gym.make("Walker2d-v4").spec.reward_threshold
>>> 

Code example

No response

System info

No response

Additional context

No response

Checklist

pseudo-rnd-thoughts commented 2 months ago

Reward threshold are not used by anthing so this shouldn't effect anything.

modanesh commented 2 months ago

I need it for my research, and that's why I created this issue. For some other envs, there are reward thresholds.

pseudo-rnd-thoughts commented 2 months ago

@Kallinteris-Andreas does a reward threshold exist? Otherwise, I don't think there is an upper bound From the looks of it, agents can achieve around 3500 mean reward though this might not be max

Kallinteris-Andreas commented 2 months ago

From what I can understand reward threshold is being used to indicate that the environment is solved when the episodic is return is reached.

I am not sure how the reward threshold values were given. In particular the Reward thresholds for MuJoCo environments is: Reacher -3.75, Seems about right. Pusher 0, unreachble / Impossible. Inverted pendulum 950, should be 1000 Inverted double pendulum 9100, Should be 9500 Half cheetah 4800, Should be around 12K Hopper 3800, swimmer 360, Is correct Ant 6000, While hard it is possible to get higher.

The other MuJoCo environments do not have A reward threshold

modanesh commented 2 months ago

Thanks @Kallinteris-Andreas and @pseudo-rnd-thoughts