Erik Frey

quat_to_euler can produce NaNs

1

`jp.arcsin` can produce a NaN in `quat_to_euler` under certain conditions, which suggests that `jp.norm(q)` > 1 + epsilon. This is probably a numerical issue tied to XLA - the bug...

TPU/GPU visualization for learning from pixels

6

Brax's rendering is CPU-only at the moment. For agents with vision, it would be useful for rendering to happen on accelerator so that training can stay fast. [Vispy](https://vispy.org/) may be...

MuJoCo converter example colab

2

@benelot proposed in #68 that a colab showing off the MuJoCo converter would help more users discover that Brax has this capability. Benjamin, if you build such a colab, we...

Mesh colliders

7

More "real" simulations often involve complex collider geometry that must be modeled accurately in the simulator. Convex hull is probably sufficient. Would need to support not just MeshMesh, but also...

Make accelerator-friendly versions of a subset of the gym wrappers

13

Candidates are: gym.wrappers.RecordEpisodeStatistics gym.wrappers.ClipAction gym.wrappers.NormalizeObservation gym.wrappers.TransformObservation gym.wrappers.NormalizeReward gym.wrappers.TransformReward And also possibly a new wrapper: gym.wrappers.RecordHTML /cc @vwxyzjn @jkterry1

Reduced coordinates / featherstone for joints

1

Use something like the [Featherstone algorithm](https://en.wikipedia.org/wiki/Featherstone%27s_algorithm) to calculate joint state in reduced coordinates - one coordinate for each rotational/translational degree of freedom for each joint. This would be a major...

Make it easier to develop local changes to js

5

`html.io` currently refers to a static cdn for the js, which is awkward for trying out changes to the js itself. We have a hacky example of local dev of...

* Add examples to docstrings * Make sure args / returns are fully fleshed out * Add documentation to config format, in particular to joint config which can be confusing...

Lower jit times

3

Jit times for training environments can be quite high, up to 6-7 minutes for free colabs communicating with TPUs. Some options here are: - Code reduction and consolidation lowers tracing...

Brax has a broad-phase culling algorithm that works pretty well on GPU: the [Brax Multi-Agent](https://colab.research.google.com/github/google/brax/blob/main/notebooks/multiagent.ipynb) colab demonstrates a scene with many bodies all interacting. One challenge with broad-phase is that...