Shaping On/Off Thruster Error into the Wrench Nullspace for Real-Time Control

Many robots, spacecraft especially, act on the world through valves that are either open or closed. A cold-gas thruster fires or it doesn't. The optimal-control solvers that now run within milliseconds all assume smooth, continuous commands, so their output has to be rounded to on/off at the last moment. That rounding injects an unplanned force and torque (together, the "wrench") on every control step. Communications engineers solved a closely matching problem for antenna arrays driven by 1-bit converters. Their solution is spatial noise shaping: the rounding error of one antenna is fed forward into the next, so that the combined error ends up in directions where nobody is listening. We propose to carry that idea over to actuator arrays. For a robot with more thrusters than it has directions of motion, there is a built-in "direction where nobody is listening": the nullspace of the thruster allocation matrix, meaning the combinations of firings that produce zero net wrench. The goal is a quantizer that costs microseconds and runs after any real-time solver, pushes most rounding error into that nullspace, and gives a bounded disturbance that a model predictive controller can plan around with a stability guarantee.

The work matters because the two current options are both poor. One is to round each thruster independently, which ignores the coupling between them. The other is to solve the exact on/off problem as a mixed-integer program, a combinatorial search whose cost grows exponentially with the number of thrusters. The intermediate idea has worked well in arrays: on a 256-antenna array, spatial shaping cut beampattern error from about 75 dB to about 48 dB compared with independent rounding (One-Bit Sigma-Delta DFRC Waveform Design). Nobody has asked what the same structure buys inside a feedback loop with a moving plant and a hard deadline.

Core Claim

Let $B \in \mathbb{R}^{6 \times M}$ map $M$ thruster firings to a six-dimensional body wrench. Let $u_c \in [0,1]^M$ be the continuous command a model predictive controller (a controller that re-solves a short-horizon optimal control problem at every step, called MPC from here on) produces. Let $b \in \{0,1\}^M$ be what the valves actually do. We quantize thrusters one at a time. Each thruster's input is its continuous command plus a weighted sum of the rounding errors $e_j$ already committed by earlier thrusters. The weights come from a strictly lower-triangular feedback matrix $F$, so $b = Q(u_c + Fe)$, where $Q$ rounds to $\{0,1\}$, and $u_c - b = (I-F)e$. Choosing $F$ from the Cholesky-type factorization $B^\top B + \lambda W = (I-F)^{-\top} D (I-F)^{-1}$ makes the wrench error

$$\|B(u_c - b)\|^2 + \lambda\, e^\top (I-F)^\top W (I-F) e \;=\; \sum_{m=1}^{M} d_m\, e_m^2 .$$

Here $D = \mathrm{diag}(d_m)$ holds the pivots of the factorization, $W$ is a positive diagonal fuel-weighting matrix, and $\lambda \ge 0$ trades wrench accuracy against wasted fuel. When $\lambda = 0$ and $M > 6$, $B^\top B$ has rank 6. That leaves $M-6$ zero pivots, and the rounding errors of those thrusters are absorbed by the nullspace at no wrench cost. This is the actuator-space analogue of the spatial noise notch. Mathematically it is the decision-feedback, or nearest-plane, structure. The difference is that it is applied after a real-time solver and analyzed as a closed-loop disturbance. We add a one-step temporal error term on top, so that residual wrench error is also pushed to high frequencies, where the plant's inertia filters it out.

The second half of the claim deals with stability. Spatial Sigma-Delta modulators, which shape quantization error by feeding it forward, have a known failure called overload. If a modulator input leaves the quantizer's range, its error states grow without bound (One-Bit Sigma-Delta MIMO Precoding). Here overload is avoided if each $u_{c,m}$ stays within a margin $\rho_m = \tfrac12 \sum_j |F_{mj}|$ of the bounds $0$ and $1$. That condition is just a tightened box constraint, which any MPC solver accepts at no extra cost. Under it, $|e_m| \le \tfrac12$, so the injected wrench is bounded by $\tfrac12\|B(I-F)\|_{\infty}$. A bounded disturbance is exactly what tube-based robust MPC and input-to-state stability arguments need. The proof therefore requires boundedness, not the independent-noise assumption that limit cycles break.

Research Program

The program has four parts.

The first is the quantizer. We build it for general allocation matrices, with an ordering heuristic: pivoting the factorization so that high-authority thrusters are quantized first. We also handle minimum-impulse and minimum-dwell constraints by restricting which thrusters may switch at each step.

The second is the closed-loop analysis. We couple the bounded shaped disturbance to the plant through a robust-MPC tube. This contrasts with Real-Time Distributed MPC with Limited Communication Data Rates, which proves stability by shrinking a multi-bit quantization interval over time with a cyclic small-gain argument. That route is unavailable on 1-bit valves, whose interval cannot shrink, so our margin comes from the nullspace instead. We will state explicitly when our bound is looser than theirs.

The third is solver integration. The tightened box is passed to ProxDDP (a constrained trajectory solver that uses proximal and augmented-Lagrangian updates) through ALIGATOR (ProxDDP), and to GuSTO (sequential convex programming with convergence guarantees) through GuSTO.jl (GuSTO). Both solvers stay unchanged; only the actuator bounds move.

The fourth is a sanity check that ties the work back to its source. For a uniform linear antenna steering model, the construction should reduce to the first-order spatial Sigma-Delta modulator of One-Bit Sigma-Delta MIMO Precoding. We will verify this on their setup.

The test platforms are a free-flyer with redundant on/off cold-gas thrusters (12 to 16 nozzles), simulated in pyastrobee (Deformable Cargo Transport in Microgravity with Astrobee) with the actuation layer replaced by binary valves, and the 12-dimensional Astrobee cluttered planning tasks from GuSTO. To test scaling, we use synthetic dense arrays with $M$ from 32 to 256 small binary actuators, a regime where exact search is out of the question.

Baselines

The naive reference is independent rounding or per-thruster pulse-width modulation of the MPC output. This is what current pipelines do after continuous solvers, according to the research file's reading of the GuSTO, ProxDDP and Astrobee work. It shares the same solver, horizon and time budget, and it shows what doing nothing costs.

The second baseline adds temporal shaping only. Each thruster runs its own first-order Sigma-Delta loop over time, carrying its own rounding error to the next step. This is the strongest cheap competitor, because the plant's inertia already filters high-frequency error. It isolates whether coupling across thrusters adds anything beyond coupling across time.

The strongest baseline is exact per-step allocation plus temporal error feedback. At each step, we find the binary vector that minimizes the wrench error of the error-corrected command, by exhaustive search for $M \le 16$ and by a mixed-integer solver under a fixed wall-clock cap for larger $M$. For small arrays this is the optimum our quantizer approximates. Exact combinatorial search is known to be impractical at array scale in the antenna setting (One-Bit Sigma-Delta DFRC Waveform Design). We found no reported result on Lacuna for tracking error under on/off thruster allocation on any of these benchmarks, so we will establish the reference numbers ourselves. The closest reported figures are the 27 dB array-error gap above and ProxDDP's under 6 ms per MPC update on the 12-joint Solo quadruped, which sets the deadline scale.

Two ablations isolate the key ingredients. The first sets $\lambda$ large, which removes the nullspace absorption and leaves only error ordering, to test whether the gains come from the nullspace. The second removes the box tightening, to test whether overload margins are needed for stability or are merely conservative. All methods share identical MPC tuning, horizons, disturbance seeds and CPU budgets, and every quantizer is timed inside the control period. Beating the strongest baseline means one of two things. For $M \le 16$, it means matching its tracking error within 10% at at least ten times less compute. For $M \ge 64$, it means lower tracking error than the time-capped mixed-integer solver, at equal fuel.

Why not just…?

Why not just use temporal Sigma-Delta per thruster, since the plant filters high frequencies anyway? This is the real threat, and we take it seriously. If the control rate is high relative to the plant bandwidth, temporal shaping may already be enough, and spatial coupling would add nothing. Our prediction is that the spatial benefit grows with the redundancy ratio $M/6$ and with tighter minimum-dwell limits, both of which reduce the effective oversampling in time. The benefit should vanish when $M = 6$, where there is no nullspace. We will sweep both factors. If temporal shaping matches us everywhere, we will report that the spatial ingredient is unnecessary for thruster control.

Why not just solve the exact binary allocation per step? For 12 thrusters, exhaustive search over 4,096 patterns fits in the deadline. We expect it to be as good or slightly better than our quantizer there. Our case rests on two other things. One is large arrays, where its cost grows exponentially while ours grows as $O(M^2)$. The other is the overload condition, which gives a guaranteed disturbance bound to plan against; per-step exact search offers no comparable a priori bound once temporal error is carried over. The $M$ sweep settles where the crossover lies.

Why not just round, and let robust MPC treat the error as a bounded disturbance? It can. But the disturbance bound for independent rounding scales with $\|B\|$ over all thrusters, so the robust tube becomes wide and the controller becomes conservative. Shaping shrinks the bound by the zero pivots. Comparing tube widths and obstacle-course success rates under both bounds answers this directly.

Why not just feed rounding error into the next MPC solve as a known disturbance? Doing so corrects error one step late at full solve cost. It is complementary, and it will be included as an additional arm, combined with each quantizer.

Proposed Experiments

The decisive check is the redundancy-by-dwell sweep on the pyastrobee free-flyer. We measure RMS position and attitude tracking error, fuel used, valve switching rate, and the quantizer's share of the control period. If nullspace shaping beats temporal-only shaping only when $M/6 > 1$ and dwell limits bind, the mechanism is confirmed. If it beats it uniformly, something else, likely the ordering, is doing the work, and the $\lambda$ ablation will show it.

The second check runs the GuSTO cluttered benchmark with binary actuation. We report success rate against GuSTO's 97% with continuous actuation (GuSTO), to see how much of the continuous-actuation performance survives quantization. We also count constraint violations with and without box tightening, which tests whether overload is a real failure in closed loop.

The third check is the dense-array scaling study, $M$ from 32 to 256, against the time-capped mixed-integer solver.

The fourth check probes the known failure mode. We hold the vehicle at a steady hover, a constant command, which is where limit cycles appear in antenna arrays (One-Bit Sigma-Delta MIMO Precoding). We measure whether valve chattering and fuel waste rise, and whether small dithering fixes this at acceptable cost.

The last check is a fidelity test. In the antenna setting, closely spaced elements couple electromagnetically (One-Bit Sigma-Delta DFRC Waveform Design). The thruster analogue is plume interaction and allocation-matrix error. We perturb $B$ by 5 to 20% to measure how fast the nullspace gain degrades.