The feature, motivation and pitch

Dear MuJoCo-team,

thank you for providing and maintaining such a great simulator. Currently, I am working as an SDE intern at Amazon robotics, using MuJoCo for large-scale simulations. Within our internal software stack, we are using the rollout() function for large-scale batched workflows.

Problem:
In the early stages of our simulations, the computed trajectories are often unstable, occasionally leading to computationally infeasible states (e.g., singular Hessian) in one of the batches, resulting in an exception. Consequently, when the simulation crashes, all other batches are lost.

Potential Solution:
To address this issue, could the rollout() function be adjusted to handle exceptions in a similar way as warnings? Specifically, when an exception occurs in a batch, the remainder of that trajectory could be filled with the current state and sensor values.

Thank you very much in advance, looking forward to the discussion.

Best,
Florian

Alternatives

No response

Additional context

No response