Racing Drone

What Racing Drone is

Racing Drone is the mission repo of the Branes.AI platform: the whole organism. Cortex perceives the world and plans where to fly, Reflex makes the airframe follow that path and keeps it safe, and Synapse runs the learned models. This repo owns the mission, wires them to a vehicle, and measures the result.

An autonomous racing drone is one of the hardest real-time embodied-AI problems that fits on a desk. Every weakness in perception latency, state-estimation accuracy, control bandwidth or compute efficiency shows up as lap time or as a crash, which makes racing the platform’s forcing function (why racing).

Baselines, not one stack

A racing drone is a baseline: components composed into fixed slots (estimator, planner, controller, safety filter, …). Every baseline builds into the same SITL, batch, replay and flight-stack executables, so baselines compare like for like.

Deterministic and replayable

Integer-nanosecond time and a log of every input and output. A SITL run replays record for record, and CI checks that logs are byte-identical.

Scored like a race

Laps, gate errors, collisions and success rate over many seeds, on an A2RL-class vehicle and course. Every phase ends in a gate a baseline must pass.

Mixed-precision by design

The stack enables quantization in each operator. A precision sweep flies the same baseline in IEEE float, custom float, posit, fixpnt and more, to find what precision each component needs. Mixed-precision is the ultimate energy safer.

The platform at a glance

Repo	Analogue	What it provides to the racing drone	Time scale
cortex	Cerebral cortex	Perception (VIO, gate detection, SLAM), world model, path planning	~5–100 Hz
synapse	The synapses	Learned models as static, allocation-free compute graphs	Deterministic
reflex	Brainstem, spinal cord	Stabilization, trajectory tracking (PID → MPC), fault recovery, safety envelopes	~100 Hz–8 kHz
racing-drone	The whole organism	Mission logic, vehicle integration, simulation and flight testing	—

Where to go next

• Introduction: the vision, the platform, and what this repo owns.
• Build & Test: build, run a closed loop, score it, replay it.
• Autonomous drone racing: why racing, from the IROS competitions to A2RL.
• Mission requirements: the mission profile, the reference vehicle, and the requirements.
• Framework architecture: slots, data contracts, baselines and data flow.
• Implementation plan: the phases and their baselines.
• Documentation map: every document, by folder.

Status: Phase 0 (the framework skeleton) is complete, and Phase 1 has started. SITL closes the loop on the reflex dynamics model, and the first baseline, B0 gt-pid, flies the course: gate P1 runs 10 seeds on every PR and 500 seeds nightly.