Warm-Start Portfolios: Keeping Several Solution Basins Alive in Real-Time Trajectory Optimization

Robots that plan their motion by solving an optimization problem every few milliseconds depend on a good first guess. Trajectory optimizers are local: started near one way of going around an obstacle, they refine that way and never discover the other. The standard fix for the latency problem is to train a neural network offline to predict a near-optimal initial trajectory (a "warm start") so the solver needs fewer iterations. These predictors output one guess. When the problem has several distinct good solutions, such as passing an obstacle on the left or on the right, or yielding to a car or going ahead of it, a single-output regressor either averages the modes into an infeasible guess or commits to one mode early. This is the same failure that reinforcement learning researchers call premature convergence. A learned policy concentrates on whatever worked first and stops visiting the alternatives. Here the "policy" is the warm-start network, and the cost of collapse is concrete. On the International Space Station, a learned warm start for the Astrobee robot did worse than a straight-line cold start when it met unfamiliar obstacles, because the solver spent its iterations undoing the bad guess (Deep Learning Warm Starts for Trajectory Optimization on the International Space Station).

We propose to replace the single warm start with a small portfolio. A network with K output heads proposes K initial trajectories. It is trained so that the heads land in different solution basins and so that the best of the K, after a fixed and shared iteration budget, has low cost. All K are refined in parallel within the control deadline, and the controller executes the best one. The claim is that multimodality in the warm start, if trained against the right objective, buys robustness to mode errors and to unfamiliar inputs at nearly constant latency. The right objective is the cost the solver actually reaches by the deadline, not the regression error to a single label. Getting there requires borrowing anti-collapse mechanisms from exploration research.

Core Claim

A warm-start predictor fails in multimodal problems for the same structural reason a policy-gradient agent collapses. Its training objective rewards concentrating on one answer. We make two changes that are each simple but have not been combined in the real-time setting.

The first change is to train the K heads with a best-of-K loss plus an exclusion term. Only the head whose post-solve trajectory is best is pulled toward that basin's solution. The other heads are discouraged from occupying the same basin. This is a direct transfer of the competitive-exclusion idea in Diversity Through Exclusion. In that paper, sub-policy heads receive zeroed reward where another head already found reward, which keeps a multi-head agent from collapsing onto one attractor. We apply it to proposal heads rather than policy heads.

The second change is to make the loss budget-aware. We score each head by the cost reached after exactly $n$ solver iterations, where $Kn$ equals the iterations a single warm start would get. The network therefore learns how much diversity a given deadline can afford. With mode labels $\{\tau^\star_m\}$ from offline enumeration and head outputs $\hat\tau_k$, the training loss is

$$ \mathcal{L} = \min_k J\big(S_n(\hat\tau_k)\big) \;+\; \lambda \sum_{k \neq k'} \max\big(0,\; \delta - d(\hat\tau_k, \hat\tau_{k'})\big). $$

Here $S_n$ is $n$ iterations of the online solver, $J$ is trajectory cost plus a constraint-violation penalty, $d$ is a distance between trajectories, and $\delta$ is a minimum separation that keeps heads from duplicating each other. Where $S_n$ is not differentiable, the first term is replaced by regression of the winning head to its nearest mode label.

Relation to Prior Work

Three lines of work motivate the gap. Learned warm starters cut iterations sharply but are unimodal. The Astrobee work reduced iterations by 60% for rotations and 50% for obstacle avoidance and failed out of distribution (ISS warm starts). Regression-based warm starts for catching flying objects predict a single B-spline (Trajectory planning for optimal robot catching in real-time). An unrolled first-order solver trained on fixed-point residuals took a vehicle problem from more than 600 iterations to 48 (End-to-End Learning to Warm-Start for Real-Time Quadratic Optimization). We reuse that last paper's training-through-the-solver technique for the budget-aware term, and add multiple heads.

The closest existing work is Learning Homotopy Prediction for Optimization-Based Trajectory Planners. A transformer there labels each obstacle "pass before, pass after, or ignore," which turns the planning problem into a convex quadratic program. It reached 95.5% success at SUMO intersections. It commits to one discrete decision and has no fallback when that decision becomes infeasible. Our portfolio keeps alternatives alive and does not require a hand-defined homotopy vocabulary. Multimodal Trajectory Optimization for Motion Planning does recover several modes, by sampling, clustering, and refining. It takes about 73 seconds per task, against 20 seconds for a single-mode optimizer. We use it, or brute-force multi-restart solving, only offline as a label generator, and distill its modes into a network that runs in milliseconds.

Work from the exploration side constrains the design. CoVO-MPC shapes a single Gaussian sampling distribution to the local cost curvature and matched 1,024-sample MPPI (sampling-based model predictive control) with 64 samples. That shaping assumes a locally quadratic cost, so it should speed collapse into the nearest basin rather than prevent it. We treat it as a strong unimodal competitor, not a building block. Energy-based policies and Diffusion MPC represent multimodality well but pay for it with iterative sampling, which the file identifies as too slow for fast control loops. A K-head network is the cheapest multimodal proposal we can think of, with one forward pass. GAGE motivates our online rule. It enforces a lower bound on exploration variance while task performance is below target. Analogously, we keep at least one head assigned to a non-incumbent basin at every control step, even when the incumbent is winning.

Research Program

We will build the system in three stages.

The first stage is offline label generation. For each training problem instance, we run many randomized restarts of the target solver, and multimodal trajectory optimization where it is feasible. We then cluster the converged solutions into basins. This yields a set of mode labels per instance and a ground-truth count of how many competitive basins exist.

The second stage trains the K-head proposer with the loss above. The first term uses the unrolled-solver variant where the online solver is a convex splitting method. It falls back to winner-takes-all regression where the solver is sequential convex programming.

The third stage is online use inside receding-horizon control. All heads are refined in parallel, batched on a GPU or vectorized on a CPU. The best feasible result is executed. On the next tick, the shifted previous winner occupies one slot, and the network refills the remaining K−1 slots with the reserved-alternative rule.

The analysis questions are the following. How does the best K trade off against the per-head iteration count n at a fixed deadline? Does the portfolio's worst case on unfamiliar inputs stay at or above cold-start quality, since a cold start can itself be one of the slots? How much of the gain comes from exclusion as opposed to simply having K outputs? We will measure latency on both a desktop GPU and an embedded-class CPU, because DiffMPC shows that parallel solvers can lose to sequential ones on embedded hardware. Parallel refinement of K heads faces the same risk.

Baselines

The naive reference is a cold start: a straight-line or zero initial guess given the full iteration budget. The second reference is a single-head learned warm start, trained the same way as in the Astrobee and unrolled-QP papers and given the full budget. The Astrobee paper reports 50–60% iteration reduction for this baseline. The unrolled-QP paper reports 48 versus more than 600 iterations on vehicle dynamics. Beating it means equal cost in distribution and strictly better cost-at-deadline and success when several basins exist or the input is unfamiliar.

On SUMO intersections, the strongest competitor is the homotopy-classification planner. It reports 95.5% overall success, 88% on high-traffic left turns, and a 4% collision rate. We will retrain it on the same expert data and give it the same wall-clock budget. A meaningful win is higher success on the left-turn subset, where homotopy errors dominate, without a higher collision rate.

For sampling-based control, CoVO-MPC at 64 samples is the competitor. CoVO-MPC reports 43–54% lower control cost than standard MPPI. We will run it at matched wall-clock time, and the portfolio must reach lower cost in scenes with disjoint corridors. All learned methods get the same training instances, the same hyperparameter search budget, and the same total solver iterations ($Kn$ fixed).

The ablations isolate the key ingredient. The first removes the exclusion term, leaving pure best-of-K training. The second keeps exclusion but drops the budget-aware term. The third, likely the most informative, replaces the heads with K random perturbations of the single-head prediction.

Why not just…?

Why not just run K randomly perturbed copies of one warm start in parallel? This is the cheapest portfolio and a real threat. On a GPU, random restarts cost almost nothing extra. Perturbations of a single prediction stay near its basin, however, so in problems where basins are far apart, such as left versus right around a large obstacle, they should rarely reach the other basin. We expect this baseline to win when basins are close and to lose when they are distant. The random-perturbation ablation, stratified by inter-basin distance from the offline labels, settles this directly. If it matches the learned heads everywhere, the learning is unnecessary and we will report that.

Why not just retrieve K diverse nearest neighbors from a solution library? Single-neighbor lookup lost to regression models in the catching work (robot catching). Nearest-neighbor warm starts also degraded under parameter shift in the unrolled-QP work (learned QP warm-start). Those papers tested one neighbor, not a diversity-selected set, so the question remains open. We will include K-neighbor retrieval with farthest-point selection among close candidates, using the same library that trains the network. We expect it to be competitive in distribution and to degrade on unfamiliar obstacle layouts, which the Astrobee split is designed to probe.

Why not just use one warm start plus a cold start in parallel (K=2)? This repairs the documented failure on unfamiliar inputs and is trivial to deploy. It does not supply a second competitive basin when the warm start picks the wrong mode in a familiar scene. Our K sweep includes this exact configuration. If K=2 with a cold-start slot captures most of the gain, the honest conclusion is that robustness, not multimodality, was the bottleneck.

Why not just take the top-K homotopy assignments from the existing classifier? This is a strong idea for driving, where discrete homotopy labels are natural, and we will run it as a SUMO baseline. It does not transfer to manipulator or free-flyer problems, which lack a clean per-obstacle label vocabulary. That transfer is why the Astrobee and grasping benchmarks are part of the evaluation.

Proposed Experiments

We use four testbeds from the file. The first is the NASA Astrobee ISS dataset, about 11,000 trajectories solved with a sequential convex programming solver, using its reported split into familiar and unfamiliar obstacles. The second is a SUMO four-way unsignalized intersection, including the high-traffic left turn, regenerated from roughly 30,000 frames of expert A* data. The third is the vehicle-dynamics QP from the learned-warm-start paper, a sanity check where there is one basin and the portfolio should not lose much. The fourth is UR3 and KUKA grasping scenes with left/right reach alternatives, taken from the multimodal trajectory optimization paper. We will regenerate the Astrobee and grasping instances if the original data are unavailable.

The metrics are success rate, cost reached at a fixed deadline, iterations to tolerance, worst-case cost ratio to cold start on unfamiliar inputs, and mode coverage. Mode coverage is the fraction of offline-identified basins hit by at least one head. Latency is reported on both GPU and CPU.

Three results would decide the question. First, on instances where offline enumeration finds two or more basins within 10% of each other in cost, the portfolio must beat both the single-head warm start and random perturbation in deadline cost. Second, on unfamiliar Astrobee obstacles, the portfolio must never fall below cold-start quality, the failure the original paper documented. Third, on the single-basin QP, the loss from splitting the budget must stay small, and that loss quantifies the price of diversity. If the exclusion term adds no coverage over pure best-of-K training, we drop the exploration framing and present the result as a budget-aware multi-hypothesis warm start.