dreamerv2 VS cleanrl

In fact, they found that just reinforce gradients work in DM control now too: Dreamerv2 GitHub (they just needed to turn off gradients through the action path - which I guess was being passed back with straight-through estimation? I'm actually having a difficult time telling how the gradient is different on the action vs policy.log_prob(action)).

This is my implementation of Hafner et al. DreamerV2 algorithm. I found the PlaNet/Dreamer/DreamerV2 paper series to be some of the coolest RL research in recent years, showing convincingly that MBRL (model-based RL) does work and is competitive with model-free algorithms. And we all know that AGI will be model-based, right? :)

Last I checked, for running off a single node, the state of the art was Dreamerv2 https://github.com/danijar/dreamerv2

PettingZoo 1.24.0 is now live! This release includes Python 3.11 support, updated Chess and Hanabi environment versions, and many bugfixes, documentation updates and testing expansions. We are also very excited to announce 3 tutorials using Stable-Baselines3, and a full training script using CleanRL with TensorBoard and WandB.

Here's where the problem starts: after implementing a custom environment that follows the typical gymnasium interface, and use a slightly adjusted PPO implementation from CleanRL, I cannot get the agent to learn anything at all, even though this specific implementation seems to work just fine on basic gymnasium examples. I am hoping the RL community here can help me with some useful pointers.

Try out a standard implementation with some standard parameters from here: https://github.com/vwxyzjn/cleanrl/tree/master/cleanrl

I am trying to run cleanrl on the `Pendulum-v1` environment. I did that by going here and changing the default `env-id` to ` parser.add_argument("--env-id", type=str, default="Pendulum-v1",

git clone https://github.com/vwxyzjn/cleanrl.git && cd cleanrl poetry install

Can someone please help me debug my implementation of SAC. Please let me know if you have any questions. I tried comparing my work with CleanRL and caught a couple of errors. However, my implementation does diverge a lot from theirs as I wanted to test my understanding.

Shameless self-plug: as far as implementation is concerned, I am working on a (hopefully) easier to understand Dreamer architecture under the CleanRL library, toward also re-implementing Director, Dreamer-v3, and and JAX variant for faster training.

Happy to share that CleanRL now has a new algorithm called Robust Policy Optimization — 5 lines of code change to PPO to get better performance in 57 out of 61 continuous action envs 🚀 (e.g., dm_control)