[Q&A] Batch renderer only render desired env_ids for memory saving

[HaoxiangYou]

Currently calling batch_renderer_camera_sensor.read(env_ids); the render will first trigger to render all envs and then return back img[env_ids].

I wonder if it is possible to hack the batch rendering, such that I can only the render the target env_ids to save memory usage.

Or is there an easy way to modify such that I can specify env_ids need to be rendering during initialization?

[duburcqa]

This is by-design, because most of the time people when to render all envs but query each of them independently. it is more efficient to do a single rendering pass in such a case.

Could you better describe what you are trying to achieve? I'm wondering if adding rendered_envs_idx option to the sensor could fit your need? This means that you won't be able to render any other env with this sensor.

[HaoxiangYou]

Thanks for your reply!

My use case is a bit unconventional. I’m experimenting with a policy-gradient setup where, for each policy action, I spawn several auxiliary environments that start from nominal (or reset) states and are only used to aggregate rewards/gradients under action perturbations. These auxiliary envs do not require RGB observations, which can significantly reduce GPU memory usage.

I appreciate your time! If you don’t have enough bandwidth to address this, I can try modifying Genesis on my side. It would be very helpful if you could point me to files or modules that are worth looking into.