Developer Joseph Suarez optimized the task to 0.14 seconds.
I trained this with pufferlib. Pufferlib is absurdly fast RL training. Like, absurdly, absurdly fast. You're often only limited by your env speed.
There's something called mujoco warp. It's a joint project between NVidia & google (judging by the commits coming from both orgs)
I just trained cartpole in mujoco at 18 million steps per second. This policy learned in **less than 3 seconds**
rollout policy batch size was 8192 agents
@yacineMTB It's 0.14 seconds with the standard pufferlib env
I just trained cartpole in mujoco at 18 million steps per second. This policy learned in **less than 3 seconds**
rollout policy batch size was 8192 agents
Developer Joseph Suarez optimized the task to 0.14 seconds.
I trained this with pufferlib. Pufferlib is absurdly fast RL training. Like, absurdly, absurdly fast. You're often only limited by your env speed.
There's something called mujoco warp. It's a joint project between NVidia & google (judging by the commits coming from both orgs)
I just trained cartpole in mujoco at 18 million steps per second. This policy learned in **less than 3 seconds**
rollout policy batch size was 8192 agents
Users are excited about Pufferlib training a Cartpole policy in MuJoCo at 18 million steps per second or in 0.14 seconds because they find the performance incredible and plan to explore the library themselves.