Elodin Open-Sources AI Racing Harness with Real-Time Betaflight Simulation for AI Grand Prix Contestants
Elodin has released an open-source simulation harness for contestants of the AI Grand Prix. The harness matches the published competition constraints and message format, giving teams a way to prepare while waiting for the official Round 1 virtual qualifier sim. It runs against real Betaflight firmware — which Dan from Elodin notes requires at least 1000 sensor samples per second to run correctly in real time.
Key Technical Details
- Direct camera sensor generation: Instead of connecting to Unreal or a similar game engine (heavy for this use case), the harness generates camera sensor samples directly inside the simulation loop. This keeps it lightweight and easier to use.
- Built with Rust Bevy game engine: Currently the visual fidelity is basic, but using Bevy should allow quick improvements. The team expects contestants to be able to shift their implementations to the official competition sim once it lands.
- Real Betaflight integration: The harness interfaces with the actual Betaflight flight controller stack, which is the standard for racing drones. The 1000-sample-per-second requirement was discovered during development.
The harness is open source and available now. It's not fancy yet, but it serves the practical purpose of letting teams test their AI agents against realistic constraints before the official sim is released. Feedback is welcome.
📖 Read the full source: HN AI Agents
👀 See Also
Configuring OpenClaw with VAST.AI GPU Rental for Unlimited Ollama Prompts
A user describes combining VAST.AI GPU rental with Ollama and OpenClaw to bypass prompt limits, but encountered configuration challenges requiring manual JSON editing.
Flash-MOE Benchmark on M5 Max: 12.99 tok/s with Qwen3.5-397B
A benchmark of the 397-billion-parameter Qwen3.5 model running locally on a MacBook Pro M5 Max with 128GB RAM achieved 12.99 tokens per second using 4-bit quantization and cache-io-split 4, three times faster than the original 48GB benchmark.
Building a Coding Agent for 8k Context: Planner/Executor Split, Token Budgeting, and Parallel Execution
A detailed breakdown of building a CLI coding agent designed around 8k token limits, using a planner/executor architecture, strict token budgeting, and parallel task execution.
Brain: A Persistent Error Memory System for Claude Code via MCP
Brain is an open-source MCP server that gives Claude Code persistent, cross-project memory for errors and solutions. It captures error context, suggests proven fixes with confidence scores, and builds a weighted synapse network connecting errors, solutions, and code modules across all projects.