Fast drone planners have a useful mathematical trick. They describe the flight path as a smooth polynomial curve (a B-spline or Bézier curve) that is pinned down by a handful of "control points," and the whole curve is guaranteed to stay inside the convex hull of those points. So if every control point sits inside an obstacle-free box, the entire continuous path is collision-free, with no gaps between sampled time steps. The guarantee, however, covers only the planned path. The real drone follows that path through a low-level controller whose gains someone tuned by hand, and on hardware the drone drifts off the plan because of motor lag, drag, and latency. Engineers patch this by inflating obstacles by a safety margin, which they also guess by hand. The margin and the controller gains are tuned together, informally, and are rarely re-checked when the hardware changes.
We propose to replace both guesses with one search. A quality-diversity algorithm called MAP-Elites builds an archive of tracking controllers. MAP-Elites does not look for a single best solution; it keeps the best solution found for each cell of a grid of behaviors. Here each controller in the archive is indexed by two quantities: how aggressively it can track a reference path, and the radius of the "tube" around the reference that the real drone stays inside. The planner then shrinks its safe corridor by the tube radius of whichever controller is flying. Because the corridor is shrunk by exactly that radius, the convex hull guarantee on the plan carries over to the actual flight. Across the reality gap, a few hardware flights recalibrate the tube radii, and the robot chooses the fastest controller whose calibrated tube still fits the corridor.
Core Claim
The convex hull guarantee is sound only for the reference path. The missing piece is a calibrated, per-controller bound on how far the real vehicle strays from that path. The claim is that this bound should be the behavioral axis of the controller search, not a constant added afterwards. Indexing an archive by tube radius and aggressiveness turns two things that are usually hand-tuned, the controller gains and the planner's inflation margin, into one quantity that can be measured. The archive also gives the online stage a structured set of fallback controllers when the real tubes turn out larger than simulated ones.
The guarantee obtained is honest about its two parts. The first part is geometric and deterministic: the reference stays inside the shrunken corridor by the convex hull property. The second part is statistical: the real deviation stays below the calibrated radius with high probability. Together they imply that the drone stays inside free space. Formally, let $\mathcal{C}$ be the safe corridor, $\mathcal{C}\ominus r$ the corridor shrunk by distance $r$, $c_i$ the reference control points, $p_{\text{ref}}(t)$ and $p_{\text{act}}(t)$ the planned and actual positions, and $r_k$ the calibrated tube radius of controller $k$. Then
$$c_i \in \mathcal{C}\ominus r_k \ \forall i \quad\text{and}\quad \sup_{t\in[0,T]}\|p_{\text{act}}(t)-p_{\text{ref}}(t)\| \le r_k \;\;\Rightarrow\;\; p_{\text{act}}(t)\in\mathcal{C}\;\;\forall t\in[0,T].$$
Positioning Against Prior Work
The closest existing system is Semi-Episodic Learning (Towards semi-episodic learning for robot damage recovery.). It builds a MAP-Elites gait archive in simulation and uses Bayesian optimization over that archive to recover a damaged hexapod on hardware. We reuse its offline-archive/online-selection structure. That system's archive axes are displacement outcomes, and its hardware trials carry no safety bound. URSA (From Tabula Rasa to Emergent Abilities: Discovering Robot Skills via Real-World Unsupervised Quality-Diversity) discovers quadruped skills with a learned safety critic inside a world model. Its safety is a neural prediction, not a geometric containment argument. Neither work uses the archive axes to make a downstream planner's guarantee sound, and that use is our contribution.
On the planning side, EGO-Planner, the B-spline kinodynamic replanner, Bubble Planner and Reachable Bézier Polytopes all certify continuous-time safety of a nominal curve through control-point containment. These works supply our planning machinery and motivate the gap. All of them assume an ideal tracker, and their margins and weights are set manually. We keep their planners unchanged and supply the margin.
Controller auto-tuning by Bayesian optimization (Automatic LQR tuning based on Gaussian process global optimization; Combining Simulations and Real-robot Experiments for Bayesian Optimization of Bipedal Gait Stabilization) constrains our design in two ways. It finds a single optimum. It is practical only to about four parameters. And with a mismatched model it ran unstable trials on hardware before converging. Our genome is a 10–16-parameter tracking controller, too large for direct hardware BO, so the expensive search stays in simulation.
Safe sim-to-real certification also constrains the design. LyEvO certifies Lyapunov ellipsoids around equilibria, at over 31 hours on up to 1,000 CPU cores for a quadrotor. Provably Safe Sim-to-Real Transfer is limited to tabular constrained decision processes. Sim-to-Lab-to-Real shields actions with neural reachability critics. None of the three handles winding spatial corridors. We instead target the specific interface between corridor planner and tracker.
Finally, the action-space benchmark of A Benchmark Comparison of Learned Control Policies for Agile Quadrotor Flight shows that collective-thrust-and-body-rate commands transfer well while single-rotor commands do not deploy at all. We therefore place the tuned controller at the collective-thrust-and-body-rate level, above the onboard rate loop.
Research Program
The genome is the gain vector of a geometric position/velocity/attitude tracking controller that outputs collective thrust and body rates. It includes feedforward scaling on reference acceleration and jerk and a drag-compensation term, giving 10–16 parameters.
Each candidate is evaluated in Flightmare, which simulates at over 200,000 steps per second. The evaluation set is a fixed library of B-spline reference trajectories drawn from EGO-Planner runs in random pillar maps. These references are played at a ladder of time-scalings, and each run uses randomized motor time constants, latency, mass and drag.
The first descriptor is aggressiveness: the largest time-scaling at which the controller remains bounded and saturation-free. The second descriptor is the tube radius at that scaling, defined as the 95th-percentile maximum deviation $\sup_t\|p_{\text{act}}-p_{\text{ref}}\|$ across randomized episodes, computed densely in time. Fitness is negative command jitter, the mean squared change in body-rate commands, because jitter is the known failure mode of transferred controllers that LyEvO reports. MAP-Elites with a Gaussian mutation operator fills a grid of roughly 20×20 cells.
Online, each elite's real tube radius is modeled as a Gaussian process whose prior mean is its simulated radius, following the archive-as-prior idea of semi-episodic learning. Each hardware flight runs one elite on a reference library scaled to that elite's speed. Motion capture records the real maximum deviation, and the model is updated.
Deployment selection follows a fixed rule. For a corridor with available slack $s$, pick the fastest elite whose upper confidence bound on real tube radius is at most $s$. Exploration starts from slow, large-tube elites and moves toward faster ones only as confidence bounds tighten. The planner then runs unchanged on $\mathcal{C}\ominus r_k$.
We also study what happens when hardware shifts. Injected latency, an added payload, and a damaged propeller are the test cases. In each, we measure how many flights are needed before the selection rule settles on a new elite.
Baselines
The naive reference point is EGO-Planner with its shipped, hand-tuned tracking gains and a hand-chosen obstacle inflation. This is what practitioners fly today. Its real-world success of around 89% in challenging scenarios is reported in the EGO-Planner paper, and that is the published number this point inherits. We run it on identical maps, references and hardware.
The strongest tuning competitor is multi-fidelity Entropy Search in the style of Rodriguez et al. It tunes a single controller by combining Flightmare and hardware trials, with real trials costing five times more. Its inflation margin is then set to the measured real tube radius of the tuned controller. Rodriguez et al. report a 35% improvement over expert tuning with 0 falls in a 4D gait-stabilization problem. No reported result found on Lacuna for this method on quadrotor corridor tracking.
For fairness, every method gets the same simulator budget, the same number of hardware flights, and the same gain bounds. For this baseline, we restrict tuning to the 4–6 gains that a GP can handle and also run it on the full vector.
A domain-randomization baseline trains one robust controller to minimize worst-case tube radius across the full randomization range. We also train a collective-thrust-and-body-rate policy with PPO under the same randomization, following Kaufmann et al.
A multi-objective baseline runs NSGA-II on aggressiveness versus tube radius with the same evaluation count. It keeps the Pareto front and uses the same online selection rule.
Ablations isolate the key ingredient. The first removes the tube-radius descriptor, indexing the archive by aggressiveness and jitter and estimating margins afterwards. The second removes online recalibration, trusting simulated tubes. The third replaces MAP-Elites with random sampling of the same number of controllers.
Beating the strongest baseline means a specific result: zero motion-capture-measured exits from true free space, together with a higher mean speed through the same corridors, under nominal and shifted hardware. The comparison is against single-controller BO with its measured margin, at equal hardware flights.
Why not just…?
Why not just tune one controller with Bayesian optimization and inflate obstacles by its measured tube? This is the most direct competitor, and we include it as the main baseline. It yields one point on the speed–margin trade-off. A single controller cannot be both the fastest option in open space and the tightest tracker in narrow gaps. When latency or payload changes, the controller's real tube can grow past its margin. The only remedies are then a full retune or slowing down. GP-BO is also documented as practical only to about four parameters (Rodriguez et al.), and it produced unstable hardware trials when its model was wrong (Marco et al.). If the shifted-hardware experiments show the single tuned controller holding zero exits at comparable speed, our archive offers little, and we will report that.
Why not just use domain randomization to get one controller robust to everything? Wide randomization makes tracking sluggish. On xArm tasks, Auto-Tuned Sim-to-Real Transfer reports 0% real success for domain randomization against 90% and 40% for simulator auto-tuning. A worst-case-robust tracker should carry a small but slow tube everywhere. Our speed comparison at matched exit rates tests this directly.
Why not just compute a Pareto front of speed versus tube radius with multi-objective optimization? This is a real threat. With two descriptors, the useful part of our archive may be close to a Pareto front. The two methods differ under reality shift. When real dynamics reorder controllers, elites that were dominated in simulation can become the best real choices, and a Pareto front has thrown those elites away. Running NSGA-II at equal budget settles the question. If its front, fed to the same calibrated selection rule, matches the archive under the latency, payload and propeller shifts, then the quality-diversity component is unnecessary. In that case the contribution narrows to tube-calibrated corridor shrinking.
Why not just use robust or tube model predictive control, which already reasons about tracking tubes? Robust predictive control is documented as overly conservative for agile tracking (An Improved Robust Model Predictive Control Strategy for Trajectory Tracking Based on Crisscross Optimization). Its tubes also come from a disturbance model that is itself a sim-to-real guess. Measured, recalibrated tubes avoid that dependence. We do not run tube MPC on hardware, so this argument rests on cited evidence rather than our own experiment.
Proposed Experiments
Simulation uses Flightmare random pillar maps at 0.4 pillars/m². On these maps the B-spline replanner reports 100% success for ideal tracking. We expect that number to fall once realistic motor lag is simulated with hand-tuned margins, and measuring this drop is our first check.
Metrics are the following: true-free-space exit rate, computed from densely sampled actual trajectories; mean speed and completion time; command jitter; hardware flights to convergence; and crashes during adaptation. As speed reference points we cite Bubble Planner's 13.7 m/s in real woods and Kaufmann et al.'s over 45 km/h for body-rate policies. These are not directly comparable targets.
Hardware experiments run in a netted motion-capture arena with foam pillar corridors. The decisive checks are three. First, simulated tube radii must predict real ones well enough that the calibrated bound holds on at least 95% of flights, as the quantile design intends. Second, under injected 20–40 ms latency and an added payload, the archive must keep zero exits while flying faster than the tuned-single-controller baseline at equal flight budget. Third, the archive must beat the NSGA-II front under shift. If any of these fails, it identifies which component (calibration, diversity, or tube indexing) is not pulling its weight.
References
- Towards semi-episodic learning for robot damage recovery
- From Tabula Rasa to Emergent Abilities: Discovering Robot Skills via Real-World Unsupervised Quality-Diversity
- EGO-Planner: An ESDF-Free Gradient-Based Local Planner for Quadrotors
- An Efficient B-Spline-Based Kinodynamic Replanning Framework for Quadrotors
- Bubble Planner: Planning High-speed Smooth Quadrotor Trajectories using Receding Corridors
- Dynamically Feasible Path Planning in Cluttered Environments via Reachable BéZier Polytopes
- Automatic LQR tuning based on Gaussian process global optimization
- Combining Simulations and Real-Robot Experiments for Bayesian Optimization of Bipedal Gait Stabilization
- LyEvO: Lyapunov-Guided Evolutionary Optimization for Safe and Robust Sim-to-Real Policy Learning
- Provably Safe Sim-to-Real Transfer
- Sim-to-Lab-to-Real: Safe RL with Shielding and Generalization Guarantees
- A Benchmark Comparison of Learned Control Policies for Agile Quadrotor Flight
- Flightmare: A Flexible Quadrotor Simulator
- Auto-Tuned Sim-to-Real Transfer
- An Improved Robust Model Predictive Control Strategy for Trajectory Tracking Based on Crisscross Optimization