Small robots such as micro-drones and compact legged platforms face a choice. They can plan with a fast but short-sighted optimizer onboard, or they can send the planning problem to a nearby server and risk the answer arriving too late. Model predictive control (MPC) is the controller at stake here: it re-solves a short trajectory optimization every few milliseconds and executes only the first step. When a remote server computes the control command itself, a delayed packet is a delayed action, and the robot has already moved. This note proposes a different division of labor. The edge server never sends commands. It sends initial guesses, called warm starts: candidate trajectories, from a longer-horizon or global search, that the onboard solver may use as its starting point. The onboard solver always keeps its own guess, which is the previous solution shifted forward by one step. It adopts an edge seed only when a cheap onboard test shows the seed is better.
The bet is that a remote answer loses its usefulness much more slowly when it serves as a starting point than when it serves as a command. If that holds, network latency spikes and dropouts stop being safety failures and become at most a temporary loss of plan quality. Separately, the edge can escape the local traps that short-horizon onboard MPC falls into, because it can afford to explore several distinct routes around obstacles.
Core Claim
We claim that splitting receding-horizon control into edge-generated, time-stamped seeds plus an onboard admission-and-polish step can do two things together. It recovers most of the benefit of a global or long-horizon planner on compute-starved robots. It also guarantees, by construction, that performance under arbitrary network delay or loss is no worse than the onboard-only controller, up to a small fixed admission overhead. The measurable prediction is a seed half-life: the delay at which an edge seed stops improving onboard cost. This half-life should be several times longer than the command half-life, the delay at which an offloaded control command causes a measurable rise in tracking error or collisions. That gap is the whole argument for the design, and it can be measured directly.
The idea sits in a gap that the two literatures leave from opposite sides. Onboard amortized optimization, exemplified by Deep Learning Warm Starts for Trajectory Optimization on the International Space Station, cut solver iterations by 50–60% on Astrobee's ARM computer. However, it solves point-to-point problems rather than closed-loop MPC. Its learned warm starts also performed worse than cold starts on unseen obstacle layouts. That result motivates our design, because an edge server can run an actual search instead of a learned guess. The same result constrains our design, because it shows a bad seed actively hurts, so seeds must be screened. Mobile-edge computing work such as Power-Delay Tradeoff in Multi-User Mobile-Edge Computing Systems and A Dynamic Service Offloading Algorithm Based on Lyapunov Optimization in Edge Computing optimizes power and delay for abstract tasks. It lets task queues grow to length $O(V)$, where $V$ is a tunable weight on power savings. That tolerance for backlog is incompatible with control commands but harmless for optional seeds. The UAV-edge papers (Evolutionary Multi-Objective Reinforcement Learning Based Trajectory Control and Task Offloading in UAV-Assisted Mobile Edge Computing, Energy Efficient Trajectory Control and Resource Allocation in Multi-UAV-assisted MEC via Deep Reinforcement Learning) plan waypoints for drones that act as servers for ground devices. They do not model flight dynamics, and they do not offload the drone's own control problem. We do the reverse on both counts.
Research Program
Onboard side. The onboard side runs sequential quadratic programming with real-time iterations in acados, as in Towards Using Fast Embedded Model Predictive Control for Human-Aware Predictive Robot Navigation. That work reached 2.47 ms per iteration with 5 humans and under 5 ms with 30, but on the large DARKO platform. We impose a much tighter iteration budget per control period to emulate small-robot compute. A reactive safety filter in the style of the 0.68 ms Safe Set Algorithm from Safe Hierarchical Navigation in Crowded Dynamic Uncertain Environments stays onboard unconditionally.
Edge side. The edge server runs a multi-start or topology-enumerating trajectory search over a horizon several times longer than the onboard one. It uses the full map rather than the local buffer that Real-time trajectory replanning for MAVs using uniform B-splines and a 3D circular buffer is limited to. The edge returns a small set of candidate trajectories representing distinct ways around obstacles. Each candidate is stamped with the predicted robot state and time it assumed.
Admission step. A seed computed for time $t_s$ arrives at time $t_a > t_s$. The robot re-anchors it by trimming the elapsed portion, splicing it to the measured state, and forward-rolling the dynamics. It then runs one real-time iteration from each surviving candidate and from the shifted previous solution. It keeps whichever candidate gives the lowest value of
$$\hat{J}(z) = J\big(\Phi_1(z;\,x(t_a))\big),$$
where $z$ is a candidate trajectory, $x(t_a)$ is the measured state at arrival, $\Phi_1$ is one onboard solver iteration, and $J$ is the MPC cost including constraint penalties. The shifted previous solution is always a candidate. As a result, the chosen trajectory is never worse than onboard-only after one iteration. The only price is $m$ extra iterations for $m$ edge candidates, and we bound $m$ by the onboard budget.
Edge scheduling. The edge scheduler must decide what to compute, not merely when to transmit. We will compare a simple periodic request policy against one that requests seeds when the onboard problem signals trouble, for example rising cost or active constraints. Wireless bandwidth is small here, since a seed is tens of spline control points rather than sensor data. The binding constraint is therefore edge server contention, which we model as a queue.
Baselines
The naive reference is onboard-only MPC under the same constrained iteration budget, warm-started from its own shifted previous solution. This is what a practitioner would deploy today. Every result is reported relative to it, and the admission rule guarantees we cannot lose to it by more than the admission overhead. We verify that bound empirically.
The strongest onboard competitor is a hierarchical onboard stack: a slow global planner running at whatever rate the constrained computer allows, feeding the fast MPC. The reference design is H-DAGap from Safe Hierarchical Navigation in Crowded Dynamic Uncertain Environments. It reports a 100% success rate with 20 agents and 97% with 50, against a roughly 23.7% collision rate for the end-to-end learned ARENA baseline. Those figures come from full-rate compute, so we will re-run the stack at matched onboard budgets. It is the strongest competitor because it already provides global reasoning without a network. It receives the same total onboard compute, the same scenarios, and the same tuning budget, which we set as an equal number of hyperparameter trials. Beating it means higher success at equal compute, or equal success at a lower onboard budget.
The strongest learned competitor is an onboard distilled MLP warm-starter trained on edge-solver outputs, following the Astrobee recipe, which reported a 50–60% reduction in solver iterations. It is trained on the same scenario distribution the edge sees and evaluated on both in-distribution and held-out obstacle layouts.
The full-offload competitor streams edge-computed MPC commands to the robot and falls back to onboard MPC when a packet is late. It runs under the identical channel and contention traces. There is no reported result found on Lacuna for offloaded MPC under wireless fading, so this baseline is new ground rather than a reproduction.
Two ablations isolate the key ingredients. The first removes the admission test and always adopts the freshest edge seed, which isolates the screening step. The second sends a single seed instead of several distinct candidates, which isolates the value of topological diversity.
Why not just…?
Why not just run everything onboard with acados and shifted warm starts? This is the real threat. On problems without competing routes around obstacles, the shifted solution is already a near-optimal seed, and we expect little gain. The evidence that it falls short elsewhere is indirect. Onboard speed has been bought by truncating the planning horizon: two-step horizons in Online Dynamic Trajectory Optimization and Control for a Quadruped Robot, local-only maps in the B-spline MAV planner, and simple constant-velocity obstacle models on DARKO. We will settle it by stratifying results by scenario type. If the gain vanishes in dense crowds and on moving-gate tracks too, the proposal fails.
Why not just distill the edge planner into an onboard network? Astrobee's learned seeds were worse than cold starts out of distribution. An edge search has no training distribution to leave. The held-out-layout experiment tests this directly. If the distilled model matches the edge on held-out layouts, the network adds nothing and we should say so.
Why not just offload the whole MPC and stream commands? Latency-tolerant scheduling in edge computing assumes delayed work is still valid, but a delayed command is not. The seed-versus-command half-life measurement is designed to show this quantitatively under the same traces.
Why not just run a slow global planner onboard? This works whenever onboard compute can sustain the global rate. We sweep onboard budget to find the crossover rather than assume it away.
Proposed Experiments
Drone racing with moving gates. Using the AirSim Drone Racing Lab, we perturb gates online and stratify tracks by the lab's curvature metric. The edge runs a long-horizon progress-based solver inspired by Time-Optimal Planning for Quadrotor Waypoint Flight. That paper reported a 24.11 s trajectory on the NeurIPS track versus 30.11 s for the winning team, but required 40–65 minutes to optimize offline. We therefore expect only a truncated online version and report lap time relative to that offline bound.
Crowd navigation. We reproduce the 20- and 50-agent scenarios from H-DAGap and measure success, collision rate, and onboard compute per cycle.
Network conditions. Channels follow the fading models used by Mao et al. Edge contention replays load from the Shanghai Telecom traces used for LDSO.
Decisive checks. Three results decide the question. The first is the seed half-life versus command half-life curve, measured under injected fixed delays from 0 to several hundred milliseconds. The second is success rate versus onboard iteration budget, compared against onboard H-DAGap. The third is held-out obstacle performance against the distilled warm-starter. The proposal is confirmed only if seed half-life clearly exceeds command half-life and a regime exists where edge seeds beat both onboard alternatives at equal onboard compute.
References
- Deep Learning Warm Starts for Trajectory Optimization on the International Space Station
- Power-Delay Tradeoff in Multi-User Mobile-Edge Computing Systems
- A Dynamic Service Offloading Algorithm Based on Lyapunov Optimization in Edge Computing
- Evolutionary Multi-Objective Reinforcement Learning Based Trajectory Control and Task Offloading in UAV-Assisted Mobile Edge Computing
- Energy Efficient Trajectory Control and Resource Allocation in Multi-UAV-assisted MEC via Deep Reinforcement Learning
- Towards Using Fast Embedded Model Predictive Control for Human-Aware Predictive Robot Navigation
- Safe Hierarchical Navigation in Crowded Dynamic Uncertain Environments
- Real-time trajectory replanning for MAVs using uniform B-splines and a 3D circular buffer
- Online Dynamic Trajectory Optimization and Control for a Quadruped Robot
- AirSim Drone Racing Lab
- Time-optimal planning for quadrotor waypoint flight