Copper Goes End to End
Over the past few weeks, we have been expanding Copper with end-to-end autonomy examples across both aerial and ground robotics.
First, we integrated ViTFly, an end-to-end vision model, into our open-source copper-rsflight controller example. Give the system a goal a few hundred meters away, and it learns to fly toward it at high speed while avoiding obstacles.
The objective is to go way beyound publishing another demo. We want to provide robotics teams with a complete blueprint they can study, modify, and use as a starting point for their own platforms.
At the same time, we have started an MVP with a customer integrating an end-to-end model into a self-driving application.
Under the hood, the flight demo runs the ViTFly model with Hugging Face’s Candle 🤗, a lightweight machine-learning framework written in Rust. Candle can execute models on regular CPUs as well as NVIDIA GPUs through CUDA and Apple GPUs through Metal, which makes the same Rust integration portable from a developer laptop to embedded GPU platforms. The self-driving demo uses an NVIDIA deployment path based on TensorRT, building on Copper’s existing CUDA support: GPU resources are represented as CuHandles and passed directly between tasks, avoiding unnecessary transfers through CPU memory. In both cases, inference becomes part of the same observable and replayable Copper dataflow as the sensors, planning, and vehicle-control components.
Step by step, the Copper ecosystem is growing to support both classical robotics algorithms and learned policies, with the same deterministic execution, observability, and replay capabilities.
Whether you are a student looking to learn a state-of-the-art robotics stack, an aerospace engineer building flying platforms that behave like real robots, or a team working through the safety challenges of physical AI, come talk to us on Discord.
