search

makeecat / Peng

A minimal quadrotor autonomy framework in Rust (Mac, Linux, Windows)
https://makeecat.github.io/Peng/
Apache License 2.0
479 stars 22 forks source link

Request for you to check out dora-rs #22

Open magi-1 opened 4 days ago

magi-1 commented 4 days ago

dora-rs is a rust alternative to ROS2 and is under active development. I am excited about the idea of doing HITL simulation with Peng via dora-rs. When you have the time, I highly recommend you check out website.

PX4 + ROS is a great option for rapidly building a functioning drone, but it's non-trivial to add CV features to ROS. Peng and dora are both super lightweight yet (near) fully featured. Being able to use a combination of cargo, maturin, and is a dream! Whenever the zero copy GPU feature is added to dora-rs, it will rival isaac_nitros which is EXTREMLY opinionated and restrictive.

I would love to chat with the both of you and discuss the future of this stack. I will be a significant contributor if we can align on direction.

@makeecat @haixuanTao

makeecat commented 4 days ago

Hi @magi-1, thank you for your interest and suggestions! @haixuanTao and I know each other. We are working on building a rust-for-robotics community. Discord link is here: https://t.co/L5agLityaT

I am tracking dora-rs development constantly, and I am happy to discuss peng+dora option with you. We can have a chat on discord!

haixuanTao commented 2 days ago

@magi-1 Following up on this issue, I have opened a PR for dora CUDA 0 copy: https://github.com/dora-rs/dora/pull/722/files

It should be fairly easy to use if you checkout the example in the PR:

For sending:

from dora.cuda import torch_to_buffer

# ...

            buffer, metadata = torch_to_buffer(torch_tensor)
            metadata["time"] = t_send
            metadata["device"] = "cuda"
            node.send_output("latency", buffer, metadata)

For receiving:

from dora.cuda import buffer_to_ipc_handle, cudabuffer_to_torch

# ..

            # storage needs to be spawned in the same file as where it's used. Don't ask me why.
            ipc_handle = buffer_to_ipc_handle(event["value"])
            cudabuffer = ctx.open_ipc_buffer(ipc_handle)
            torch_tensor = cudabuffer_to_torch(cudabuffer, event["metadata"])

Feel free guys to leave a comment or a review :)