ADR 0007: outdoor obstacle avoidance lives in the platform and its own mission repo, not in racing-drone

Proposed · 2026-10-28 · racing-drone/docs/adr/0007-outdoor-obstacle-avoidance-placement.md

Status: Proposed Date: 2026-10-28 Related: ADR 0001, ADR 0005; requirements §2, SYS-03, DAA-09; the Branes.AI tiers in the introduction

Context

Loquercio, Kaufmann et al., “Learning high-speed flight in the wild” [1], flies a quadrotor through unknown, unmapped, static clutter (dense forest, snow-covered terrain, a collapsed building, a derailed train) with onboard sensing and compute:

• Sensing: a depth image from stereo matching, plus inertial measurements.
• Policy: a CNN maps them directly to a set of short-horizon, dynamically feasible polynomial trajectories, each with a predicted cost, in a receding-horizon loop. There is no mapping and no separate planning stage.
• Training: in simulation only, by privileged learning. It imitates a sampling-based expert (Metropolis-Hastings) that sees the full 3D point cloud and the true state. Simulated sensor noise gives zero-shot transfer to the real world.
• Results: the policy runs onboard on a Jetson TX2 [1].
  • Real world: 56 runs in natural and human-made environments never seen in training. The paper reports these success rates by speed in a figure only [1].
  • Simulation: in a dense simulated forest, 100 % success up to 5 m/s and 60 % at 10 m/s. Across all the simulated environments, the average is 70 % at 10 m/s, where no baseline completes every run even at 5 m/s [1].

The question is where to develop these techniques: here, or elsewhere.

Three facts decide it:

1. The problem is not this repo’s DAA. Our DAA (ADR 0005) detects and tracks other drones with one camera on a known course. [1] avoids static obstacles in an unknown environment, from stereo depth, while flying toward a goal.
2. It is a different mission. The task is reaching a goal through terrain the vehicle has never seen, not a lap of a known gate course. Its score is success rate against speed, not lap time. Its sensors include stereo, which the A2RL rules exclude (requirements §2, SYS-03).
3. The tiers already say where the pieces go. Algorithms live in the platform repos: perception and planning in cortex, learned-model execution in synapse, control and safety in reflex. Each mission repo owns one vehicle and mission and composes those components. The README says it directly: “algorithms do not live here”.

There is a second question underneath. The composition framework is mission-neutral in intent, but it lives in this repo as branes::racing: the slot concepts, data contracts, Stack, harnesses, log and replay, TimeBase, and the SITL core. A second mission repo would want all of it.

Decision

1. The algorithms go in the platform repos, including the offline ones:

   Piece	Repo
   Stereo depth, and the learned receding-horizon planner	cortex
   The privileged expert (the sampling-based planner the policy imitates) and the imitation-training pipeline, as offline tooling beside the planner they train	cortex
   Running the policy network as a static, allocation-free graph	synapse
   A safety filter under the learned planner (the CBF of ADR 0005, applied to obstacles)	reflex
   Simulated depth with stereo noise, as a sensor model of the simulator core	moves with the framework (Decision 4)

   The planner is not yet a Planner slot implementation. The current slot takes only the course map and the state estimate (slots.hpp). A depth-driven planner also needs a depth observation and a goal. Putting it behind the slot therefore needs a revised Planner contract, decided where the planner enters a mission (Decision 3), under the slot registry of ADR 0001.

2. The mission goes in its own mission repo, with only mission content and no algorithms, as the tiers require. That means the randomized outdoor scenes and scenarios, the vehicle and its sensor configuration, the trained weights for that vehicle, and the success-rate-against-speed scoring and acceptance. That repo is scout-drone (named after this ADR was proposed).

3. Racing-drone takes the planner only as a comparison baseline, later. Once the planner exists in cortex and the Planner contract carries a depth observation and a goal (Decision 1), a baseline may put it behind the slot and compare it with the waypoint and MPCC planners on an obstacle course. A2RL announced that Season 2 would feature “a more technically complex track featuring dynamic obstacles and mission-style tasks” [2]. Such a baseline adds a depth sensor through the existing sensor interfaces (requirements §2, DAA-09). It is therefore not an A2RL-legal baseline, and it is reported as such.

4. The framework is lifted out before a second mission repo is built on it, not copied.

   • Until then: keep mission-neutral code free of racing-only types.
   • Racing-only: gates, courses, the oracle gate projector, the scorer.
   • Mission-neutral: slots, contracts, Stack, harness, log, time base, the simulator core.
   • Where the framework goes is decided in ADR 0008: the anatomy repo.

Consequences

• Racing-drone’s scope stays the race. It doesn’t grow outdoor scenes, a privileged expert, or a training pipeline.
• The Planner contract will change before any depth-driven planner is a baseline here. The existing planners (NullPlanner, WaypointPlanner) must then be adapted to the revised signature.
• Each platform piece is reusable. The planner, the network and the filter serve any mission, including a later racing obstacle-course baseline, behind the same slot.
• The cost is coordination. The outdoor work spans three platform repos and one mission repo, and it needs the framework decision before the mission repo can use slots, harnesses and replay without copying them.
• The rule in Decision 4 constrains racing-drone now, at little cost: racing-specific types stay out of the mission-neutral headers. That boundary is what makes the later lift a move, not a rewrite.

Current state: proposed.

Sources

1. Loquercio, Kaufmann, Ranftl, Müller, Koltun, Scaramuzza, “Learning high-speed flight in the wild,” Science Robotics 6(59), abg5810 (2021), doi:10.1126/scirobotics.abg5810. Accepted version: https://rpg.ifi.uzh.ch/docs/Loquercio21_Science.pdf (accessed 2026-10-28). Code: https://github.com/uzh-rpg/agile_autonomy.
2. A2RL, “A2RL Drone Racing Returns: Season 2 Elevates Autonomous Flight” (16 Jul 2025). https://a2rl.io/blog/8/A2RL-Drone-Racing-Returns-Season-2-Elevates-Autonomous-Flight (accessed 2026-10-28).