WorksComputer Science
Capability Composition in Heterogeneous Teams via Barrier-Guided Diffusion Sampling
Abstract
The agents in a heterogeneous team can cooperate to accomplish tasks that exceed their individual capabilities. The team trajectories for a cooperative task can be generated by a generative model, provided that the model has been trained on demonstrations of the agents cooperating. Yet, demonstrations of cooperation are costly and may not exist for tasks that the team has never performed. The conditions under which the agents can combine their capabilities, however, are physical properties of the team and are known before any demonstration is collected. We encode the conditions as constraints on the team trajectories, and we enforce the constraints while sampling from generative models trained on demonstrations of each agent alone. We prove that the proposed method can generate every team trajectory that satisfies the constraints, although the demonstrations contain no cooperation. Hence the constraints determine which team trajectories can be generated, and the demonstrations determine how likely each team trajectory is. Numerical results on two cooperative tasks show the effectiveness of the method, which generates cooperative team trajectories that satisfy the constraints and requires no retraining when agents join or leave the team. Finally, we implement the method on a multi-robot testbed.
Introduction
Cooperation allows the agents in a team to accomplish tasks that exceed their individual capabilities. Cooperation in multi-agent teams has long been studied, from the allocation and scheduling of tasks [1, 2] to the control of several robots acting on a common object [3, 4]. Existing methods assume that the capabilities of the team are specified in advance, together with the conditions under which each capability can be used. The capabilities that a task requires, however, depend on the environment where the agents are deployed, and the environment is often unknown when the agents are trained. The task may then require a capability that no agent has alone, and that the agents obtain only by combining their individual capabilities. Hence the required cooperation can be neither specified nor demonstrated in advance, and generating cooperative behaviors for the task remains a challenge.
Generative models, and diffusion models in particular, are effective methods to generate the trajectories of an agent [5, 6] or of a team [7, 8], provided that the demonstrations contain the desired behavior. To overcome this limitation, we observe that the conditions under which the agents can combine their capabilities follow from the physics of the team. Hence the conditions can be written as constraints on the team trajectories without any demonstration. We then sample the team trajectories from generative models trained on demonstrations of each agent alone, and we impose the constraints during sampling using a guidance method [5, 9]. This results in a method to sample cooperative team trajectories from models pre-trained on non-cooperative demonstrations.
Related work. Task allocation and scheduling decide which agent performs which task and when [1, 2], and a trajectory planner then realizes the decision. We sample the trajectories directly from diffusion models and read the pattern of cooperation off the result, with nothing computed in advance. One line of multi-agent diffusion planners trains on demonstrations of cooperation where they exist [7, 8], and reproduces only the cooperation those demonstrations contain. Another trains on single-robot data and reconciles the resulting plans by search [10] or by projection [11]. These approaches combine constraints by intersection, as for collision avoidance. We combine them by union, because cooperation admits configurations that no agent reaches alone. Cooperation is available only while the agents are physically arranged for it, and has to be encoded as a constraint that holds at every timestep of a trajectory. Control barrier functions [12] (barriers, for short) encode such constraints. A barrier keeps a trajectory inside the set it defines. Boolean combinations of barriers give intersections and unions of the admissible sets [13, 14]. On the generative side, barriers have been used to keep diffusion samples inside a safe set [15] and to guide sampling with a constricting condition that tightens as noise falls [16]. Barriers also make flow-matching samplers safe [17, 18]. In multi-agent systems, guidance can be distributed, so that each agent computes its own Lyapunov-based guidance input [19].
Contributions. Our contributions are as follows. First, we formulate cooperative trajectory generation without demonstrations of cooperation as barrier-guided diffusion sampling, with constraints that encode the individual capabilities of the agents and the cooperative ones. Second, we propose a scheme in which the scores trained independently for each agent are summed into a product prior, and we prove that the sum is the exact score of that product at every noise level, since the densities of different agents depend on disjoint sets of variables. Third, we prove that the sampler can generate every team trajectory satisfying the constraints with nonzero probability and, when its guidance succeeds at every step, no other. Thus, every pattern of cooperation the constraints permit can be generated, and none is specified or pre-computed. Fourth, we demonstrate the method on a river crossing task, in simulation and on the Robotarium. The generated trajectories cross the river cooperatively, while sampling the prior alone yields agents that ignore each other and fail the task. We also show that the sampler requires no retraining to generate trajectories when agents join or leave the team. Finally, we consider a task in which cooperation also has to be timed. The sampler generates such trajectories without demonstrations, but less often than in the river crossing. Most failures of the sampler are consistent with local maxima of the barrier.
Problem formulation
Consider a team of agents with states and inputs , evolving in discrete time as
Let be the team state and the team input, with and . A team trajectory is the matrix , whose row block corresponds to the trajectory of agent .
Without cooperation, agent is constrained to the states and controls satisfying , where is a problem-dependent constraint function. Cooperation enlarges this set. Let describe the allowable configurations while cooperating with the rest of the team, so that the constraint set of agent is
A team trajectory is admissible when the pair lies in for every agent and at all times. Cooperation enlarges the set of admissible states and controls, while the dynamics (1) remain the same whether or not . Combining constraint sets is standard in the safe control literature [13]. Multi-robot planning combines them by intersection, since collision avoidance asks that a plan satisfy every agent’s constraints, and the feasible set may shrink [10]. In (2), instead, the second set depends on the other agents, so admits configurations that agent cannot reach alone.
A task is specified by a problem-dependent function , and a team trajectory successfully completes the task when . We are interested in tasks that no agent can complete alone and the team can complete together. This motivates the following assumption.
Assumption 2.1. (Cooperation is necessary and sufficient to complete the task) Every admissible team trajectory that completes the task also satisfies for some agent and some time . Moreover, at least one such trajectory exists.
Assumption 2.1 requires that, along every trajectory that completes the task, some agent at some timestep violates its own constraint and is admissible only through the coupling constraint . Hence no trajectory in which every agent always satisfies its own constraint completes the task, that is, cooperation is necessary. Cooperation is also sufficient, because at least one admissible trajectory completes the task.
Example 1. (Crossing a river) Consider a team of two agents, a ground vehicle and a vessel, operating in a unit square environment divided across the middle by a river of width . The task is for the vehicle to reach a location across the river. The vehicle cannot enter the water and the vessel cannot leave it, so the vehicle cannot satisfy the task alone.
Let be the position of agent , with and as its horizontal and vertical coordinates, respectively. Agent is the vehicle and agent is the vessel, with dynamics . The constraint functions are
and everywhere, since cooperation never lets the vessel leave the water. Hence when the vehicle is on land and within its own speed limit, when the vessel is on water and within its own, and when the two are within a tolerance radius and the vehicle moves within of the vessel. Figure 1 shows these sets (positions). The vehicle may therefore be in the water only while riding the vessel. While the vehicle is aboard, its speed limit applies relative to the vessel, so its velocity is the velocity of the vessel plus an input of norm at most . Since leaves , the vessel is bound by at all times, and whenever the vehicle is within of the vessel. The vehicle starts at a given position on one side of the river, the vessel at a given position on the water, and the vehicle must reach a goal across the river:
where is the tolerance on each of the three positions.
Figure 1. Positions admitted by the vehicle’s constraint set in Example 1, for one position of the vessel (square): the land, by (left), within of the vessel, by (middle), and their union (right). The disc is narrower than the river, so it reaches land only when the vessel is near it. The vehicle (circle) is aboard the vessel when and .
We study the following capability composition problem: given a task and a team of agents with constraint sets , sample team trajectories with and for every and every . We generate trajectories with a denoising diffusion model, a dynamical system that transports a Gaussian into the data distribution [20, 21]. A forward process adds noise to over diffusion time as
where is the noise schedule and is a Wiener process. We write for the density of at time , so that is the density of the demonstrations and close to the standard Gaussian . The coefficients of (3) are scalar, so noise enters each coordinate independently. Sampling reverses (3), running from back to , according to
with now a reverse-time Wiener process and the drift built from a score network trained to approximate the score . We assume that no demonstration of cooperation is available for training. Let be a set of demonstrations of agent alone, let be their density, and let be a score network trained on . The team prior is the product of the per-agent densities. We assume that each is collected in an unconstrained environment, in which agent is free to occupy any state. The demonstrations are therefore reusable across tasks, while is task-specific and enforced at sampling time. In Example 1, and contain trajectories of the vehicle and the vessel.
Main results
Generating a team trajectory through (4) requires the team score at every diffusion time, and § 2 leaves us with one score per agent. Let denote the density at time of the forward process (3) started from . The separately trained per-agent scores yield the team score exactly, as we show next.
Proposition 3.1. (Agent scores form the team score) Let be the density of at time . Then the rows of the team score for agent are its own score: for every and ,
Proof. Let be the density of at time when the forward process starts from the trajectory . It factorizes across coordinates, so , and the logarithm turns the product into a sum whose -th term depends on alone.
The identity (5) holds whenever the densities in the product depend on disjoint sets of variables. The forward process adds noise to each coordinate independently, so it smooths each variable separately. When the densities are over the same variable, as in a product of experts, smoothing the product is not the same as multiplying the smoothed densities, and the identity holds only approximately [22].
Sampling (4) with this score returns team trajectories drawn from , and there is no reason for them to be admissible or to complete the task: the per-agent densities were trained on demonstrations and carry no information about or . The constraints must therefore be imposed while the trajectory is generated. Following [16], we add a control to (4),
and choose at each step so that a scalar barrier stays non-negative along the sampled trajectory. The barrier encodes all the constraints specified in § 2. The sampler generates state trajectories only, while the constraint functions depend also on the inputs. We therefore assume that, for every agent and state , the map in (1) is injective, so that is the unique input taking to . In Example 1, . Hence the constraints along a trajectory are functions of only. Admissibility is an intersection over agents and times, and (2) is a union:
so that when the trajectory is admissible. The actuation limits do not need a term of their own, since (2) already carries each agent’s actuation limit in and the coupled regime’s in . The barrier then becomes
and the set contains exactly the admissible team trajectories that complete the task.
The sampler of [16] simulates (6) on a grid with steps ,
from to the returned trajectory , where the are independent Gaussian increments. The control uses the relaxed barrier , whose margin is continuous, nondecreasing in , vanishes at , and equals at , so that . At each step is drawn first, then is chosen so that
with and . If (9) holds at every step, then stays non-negative and for every realization of the noise [16]. In that case is admissible and completes the task, and, by Assumption 2.1, at some timestep some agent violates its own constraint and is admissible only through its coupling constraint . Let denote the state that (8) gives with and the noise already drawn, so that (9) depends on through . Linearizing at a point gives the condition
which is linear in , so the control of minimum norm that satisfies (10) is in closed form. The sampler of [16] takes the minimum-norm control with . As noted in [16], the linearization error is quadratic in the displacement , which is dominated by the noise term . The error is therefore of order , the same order as the decrease that (9) allows. Hence a control that satisfies (10) at can violate (9), and the guarantee is then lost.
To overcome this limitation, we use the sampler of [16] with three modifications. First, the relaxed barrier is evaluated on an estimate of the trajectory being formed rather than on the iterate itself, which for most of the reverse process is dominated by noise and carries little information about the trajectory the sampler returns. The estimate is the best affine estimate of from , given in Lemma 3.2 below. Second, the control is whenever the unguided update satisfies (9), instead of its linearization as in [16 Remark 2]. Third, when the unguided update violates (9), we search for locally. We first set to the control of minimum norm that satisfies (10) with , as in [16]. While (9) fails, we set , the state reached with the current control, and recompute from (10). If no control satisfying (9) is found within iterations, as happens near a local maximum of , where the gradient vanishes, we take the step with the last control. The guarantee is then lost, but the returned trajectory may still lie in , because later steps can increase again. For this reason, Algorithm 1 returns together with .
Lemma 3.2. (Affine estimate of the returned trajectory) Let be a random variable with mean and covariance . Let evolve from according to (3), and let and . Then
is given by and , so that satisfies
Moreover, is invertible for every , , and when is Gaussian.
Proof. (3) is linear with scalar coefficients, so , where is independent of and . Hence has mean and covariance , and its cross-covariance with is . The minimizer is and , which gives the expressions above. The eigenvalues of are positive, and has the same eigenvectors and the eigenvalues , so is invertible. Since and , and . When is Gaussian, is jointly Gaussian, and its conditional mean is affine and equal to (11). □
Lemma 3.2 uses only the mean and the covariance of , so is computed in closed form and does not need to be learned. During sampling, is the current iterate, and is the best affine estimate of the trajectory being formed. We therefore let and be the mean and covariance of the product prior , estimated from the training sets , and evaluate the relaxed barrier on this estimate, that is, , with margin . Theorem 4.2 of [16] is stated for the barrier , and it extends to the modified above with the same proof. The margin gives , and if (9) holds at every step, then for every , by induction. As vanishes at and , , so . In our implementation, the right-hand side of (9) is replaced by its minimum with , so that the guidance stops acting once a trajectory is well inside . The right-hand side remains non-negative, and the argument above still applies.
We next characterize the type of trajectories that the diffusion sampler can generate. In particular, we prove that any admissible team trajectory solving the task can be generated with arbitrary accuracy and nonzero probability, even when the training set does not contain such trajectories. To this end, we make the following technical assumption.
Assumption 3.3. (Monotone reachability) For every and , there exists with and a continuous path from to along which is nondecreasing.
Assumption 3.3 asks that every trajectory in be reachable from below without lowering . It fails when a trajectory in lies in a region enclosed by higher values of , so that first increases and then decreases along every path that reaches it from a lower level set. It involves only , and neither the prior nor the data. For and , denotes the open ball of radius centered at in the Frobenius norm, and denotes the support of a distribution , that is, the smallest closed set of probability one.
Algorithm 1 Barrier-guided sampling, modified from [16]
Require: per-agent scores , relaxed barrier , grid , gain , iteration limit
Draw and set the margin
for do
Draw and set
if violates (9) then
[16]
for , while (9) fails at do
end for
end if
end for
return and
Theorem 3.4. (Every trajectory in can be generated) Let Assumption 3.3 hold, let and be continuous, with the closure of , and let be the distribution of the trajectory returned by Algorithm 1. Then for every and every . If, in addition, (9) holds at every step, then .
Proof. Fix with and , and let be such that on , which exists because is continuous. Assumption 3.3 with any gives a trajectory with and a continuous path from to along which is nondecreasing. Let . At the margin is , so , and as vanishes at and ,
so the path has points with
Let , so that as , and increases strictly as decreases.
Since vanishes at it is non-negative at every step, and by (12) it increases strictly. Hence the unguided update from to satisfies (9) with strict inequality, and Algorithm 1 sets there. The recursion (8) adds to a term set by and , and , so exactly one increment carries to and returns .
The barrier is continuous and (12) is strict, so induction on gives at every step on a neighborhood of . The recursion is the unguided one there, hence continuous, and a smaller neighborhood is carried into . That neighborhood has positive probability under the Gaussian initialization and increments, so . If instead , then contains a trajectory with , because is the closure of , and the claim follows from . Finally, we prove the second claim. If (9) holds at every step, then for every realization of the noise, by the extension of [16 Theorem 4.2] given after Lemma 3.2, so that . Since is continuous, is closed, and hence . The reverse inclusion is the first claim, since every ball around a point of has positive probability.
We next ask which agents cooperate, and when. Let the times at which agent is admissible only through its coupling constraint be
and we call the coupling pattern of . The trajectories in with pattern form the set , and these sets partition . The sampler returns, with nonzero probability, the coupling pattern of every trajectory that satisfies the constraints strictly.
Corollary 3.5. (Every coupling pattern can occur) Let the hypotheses of Theorem 3.4 hold and let satisfy and, for every and , either or . Then the sampler returns a trajectory in with the coupling pattern of with nonzero probability, that is, .
Proof. The inequalities in the hypothesis are strict and , and are continuous, so there is a such that every trajectory in satisfies them and has . By (13) the coupling pattern is throughout , and Theorem 3.4 gives .
The constraint functions , and determine . By Theorem 3.4, the sampler can return every trajectory in , with probabilities determined by the demonstrations. The constraints do not prescribe a coupling pattern, because lies in whenever or . Hence contains trajectories with different coupling patterns, and Corollary 3.5 gives each of them a positive probability of being sampled. Finally, by Proposition 3.1, each agent contributes its own score to the product prior, while the constraints enter only through the guidance. Adding or removing an agent therefore changes only the constraint functions and the set of scores in the product, and no network needs to be retrained.
The probability that the sampler returns a trajectory in depends on whether (9) can be satisfied at every step, because (9) at every step guarantees that the returned trajectory lies in . Algorithm 1 searches for such a step locally, along gradients of . Because in (7) is a maximum over the two sets of each union, can have local maxima at which , and a search that reaches one may find no step satisfying (9) even when is nonempty. Controllers based on control barrier functions have the same limitation, since their quadratic programs can stall at points other than the intended ones [23]. The maxima that encode unions of sets [13] add such points.
Numerical results
We consider Example 1 over 50 timesteps, with , , and . The training sets and contain 2000 single-agent trajectories each, recorded in an environment of the same dimensions without the river. Hence the demonstrations describe how each agent moves, and contain neither the constraints nor any cooperation.1 To compute the guidance from the gradient of , we replace the min and max in by the smooth surrogates
over the entries of , with . Both surrogates are bounded above by the exact minimum and maximum, so the smoothed barrier is non-negative only on trajectories in .
The sampler generates cooperative trajectories, although neither training set contains one. No trajectory drawn from the team prior lies in (Table 1). Guided by alone, all 500 generated trajectories reach the goal, but in every one of them the vehicle enters the water without the vessel. Guided by , all 500 lie in . Figure 2 shows one trajectory of each kind. In the trajectory guided by , the vessel meets the vehicle at the river, the two cross together, and the vehicle continues to the goal. This sequence appears in neither training set.
Figure 2. River crossing of Example 1: a trajectory sampled from the prior alone, one guided by the task constraint only, and one guided by all constraints (Algorithm 1). The band is the river, open circles mark initial positions and the disc around is the goal. The vehicle’s path in the river is green where the vessel is within and amber where it is not. Right: vehicle-vessel distance along the trajectory guided by , shaded while the vessel carries the vehicle.
| river crossing | escape room | Robotarium | |
| Algorithm 1 | 100 | 30 | 100 |
| per-step projection | 100 | 36 | 99 |
| penalty guidance | 16 | 0 | 14 |
| projection after sampling | 44 | 2 | 54 |
| rejection | 0 | 0 | 0 |
Table 1. Percentage of the generated trajectories that are admissible and complete the task, averaged over five seeds.
The constraints require the two agents to cross together, and the learned prior determines where and when they do. To separate the two contributions, we replace the learned prior by a Gaussian of matched moments and keep the barrier. For each of the 500 trajectories guided by , we record the coordinate and the timestep at which the two agents come closest (Figure 3). With the Gaussian prior all trajectories also lie in , and the crossings concentrate near the point where the straight line from to meets the river. With the learned prior, the trajectories inherit the turns and pauses of the demonstrations, which are walks through two or three random waypoints at constant speed, and their crossings spread over most of the region that the horizon allows. The number of timesteps on which the two agents are within of each other varies in the same way, from 4 to 27. The constraints require only two such timesteps, so the longer couplings come from the prior.
Figure 3. 500 trajectories from the learned prior and 500 from a Gaussian prior of matched moments, both guided by . The shading counts the trajectories that cross at the coordinate and the timestep at which the two agents come closest. The red curves bound the crossings that allow reaching the goal by . The learned prior spreads its crossings over that region, and the Gaussian concentrates them near the middle.
The sampler requires no retraining when the team changes. We add a second vessel that can also carry the vehicle, so that in (2) gains a third set. Both vessels use the same score, and fixes the initial position of each, the new vessel starting on the water at distance 0.2 from . All 500 trajectories generated for this team lie in , and in 127 of them the vehicle passes from one vessel to the other. With three vehicles and three vessels, all 500 trajectories lie in . They show 144 different coupling patterns, against 12 in 100 solutions of trajectory optimization without the prior. Removing agents works in the same way, unless the remaining team cannot complete the task. With no vessel, for example, Assumption 2.1 fails and no trajectory lies in .
We implement the river crossing on the Robotarium [24], a testbed of unicycle robots in an arena of 3.2 m by 2 m. The unit square of Example 1 is mapped to a square of side 2 m in the arena, and each timestep of a trajectory is executed in 1.2 s. The robots have a physical size, so we change the constraint functions and keep the networks and the sampler. The river widens to 80 cm and grows to 32 cm, so that two robots fit within of each other, while still forces the vessel to carry the vehicle. We also add two constraints that the simulation does not need. The first requires every pair of robots to stay at least 20 cm apart, so that the robots do not collide. The second requires every robot to stay inside the arena. Finally, a constraint that holds at every timestep can still be violated between two consecutive timesteps, so we impose every constraint on the whole segment between them. We call this the Robotarium setting. Algorithm 1 places every trajectory in , with one vessel and with two. Each robot tracks its sampled positions through the Robotarium’s map from single-integrator to unicycle inputs, so the nonholonomic constraint appears only as tracking error. We execute the first 15 trajectories with one vessel and the first 5 with two. All 20 runs satisfy the constraints up to the tracking error, which stays below 5 cm. Figure 4 shows a run with two vessels, in which one vessel hands the vehicle to the other.
Figure 4. A Robotarium run with two vessels. (a) Camera frame at the handover, with executed paths of the vehicle (blue), vessel 1 (red) and vessel 2 (purple). Dashed lines are the riverbanks. (b) Distance from the vehicle to each vessel, with the time in the river shaded and the time of (a) dotted.
In the second task, an escape room, cooperation also has to be timed. Two agents start on one side of a wall with a doorway, and both have to reach an exit on the other side by the end of the horizon. The door is open at time when, at , some agent was on a pad placed before the wall, or in the doorway. The state of agent consists of its current and previous positions, and , because the door depends on the previous step. Let be the wall with the door shut, its solid part, so that is the doorway, the pad, and the signed distance to , positive outside . Then
where when is on the pad or in the doorway, so that when the door is open. The task function asks each agent to start within of its initial position and to end within of the exit. The wall is thicker than one step and the pad is more than one step from the doorway, so an agent alone cannot pass. Two agents can, one on the pad while the other enters the doorway and then holds the door. Both agents use the same score network, trained on 4000 single-agent trajectories recorded in the room without the wall.
Cooperation that has to be timed is harder to generate. In the escape room, about 30% of the trajectories generated by Algorithm 1 lie in , and in each of them one agent stands on the pad while the other enters the doorway (Figure 5). Without the local search, as in [16], no trajectory lies in . Most failures leave an agent at the shut door, which is consistent with the local maxima of discussed in § 3.
Figure 5. A trajectory of the escape room in , at four timesteps (top) and over the horizon (bottom). Dots are the agents and faint lines their last five steps. The doorway is dark when the door is shut, the disc is the pad, and the dotted line is the exit. Agent 2 is on the pad from to , so the door is open from , agent 1 enters the doorway at and holds the door, and agent 2 enters at . Bottom: timesteps on the pad (yellow), in the doorway (agent colour), and with the door open (shaded).
In all three settings, Algorithm 1 and per-step projection place far more trajectories in than the other three methods in Table 1. All methods use the same prior, barrier and noise. Per-step projection [11, 25] moves the sample toward after every reverse step with a few Newton steps, on or on , and the table reports the better of the two. Penalty guidance [5] adds to the score the gradient of a penalty on constraint violation, with a weight chosen per task. Projection after sampling moves each sample of the prior onto once sampling ends. Rejection keeps the samples of the prior that lie in . These last three methods place at most about half of the trajectories in , and almost none in the escape room. Per-step projection comes with no guarantee that the returned trajectory lies in , while Algorithm 1 has one whenever (9) holds at every step. In the river crossing and in the Robotarium setting, both methods place almost every trajectory in , and projection is faster. In the escape room, projection places 36% of the trajectories in , against 30%. It moves the samples about 400 times more than Algorithm 1, as measured by the control energy of [16], and its trajectories end farther from the prior, as measured by the maximum mean discrepancy [26]. In the river crossing, projection moves the samples about 50 times more, and its trajectories are as close to the prior as those of Algorithm 1. On images, the larger correction of per-step projection visibly degrades the samples [16].
Conclusion
In this paper we study how to generate trajectories for a team of agents, without demonstrations of cooperation, when the task requires a capability that no agent has alone. We encode the capability as constraints on the team trajectory, and guide the sampling from diffusion models trained on each agent alone with a barrier function built from the constraints. The resulting sampler can provably generate every team trajectory that satisfies the constraints. Our numerical results show that the sampler generates cooperative trajectories that satisfy the constraints, and that it requires no retraining when agents join or leave the team. On the Robotarium, the generated trajectories satisfy the constraints also when unicycle robots execute them, up to the tracking error. Existing guidance methods, however, can fail when the constraint sets are complex, for example when cooperation also has to be timed. Future work includes the design of guidance methods for complex constraint sets, and learning the coupling constraints from demonstrations of interacting agents.
Acknowledgments
This work was supported in part by NSF under award CMMI-2622263.
References
- [1]B. P. Gerkey and M. J. Matarić, “A formal analysis and taxonomy of task allocation in multi-robot systems,” International Journal of Robotics Research, vol. 23, no. 9, pp. 939–954, 2004.DOI
- [2]G. A. Korsah, A. Stentz, and M. B. Dias, “A comprehensive taxonomy for multi-robot task allocation,” International Journal of Robotics Research, vol. 32, no. 12, pp. 1495–1512, 2013.DOI
- [3]E. Tuci, M. H. M. Alkilabi, and O. Akanyeti, “Cooperative object transport in multi-robot systems: A review of the state-of-the-art,” Frontiers in Robotics and AI, vol. 5, p. 59, 2018.DOI
- [4]H. Farivarnejad and S. Berman, “Multirobot control strategies for collective transport,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 5, pp. 205–219, 2022.
- [5]M. Janner, Y. Du, J. B. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” in International Conference on Machine Learning, ser. PMLR, vol. 162, 2022.arxiv.org/abs/2205.09991
- [6]C. Chi, Z. Xu, S. Feng, E. Cousineau, Y. Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuo-motor policy learning via action diffusion,” in Robotics: Science and Systems, 2023.
- [7]Z. Zhu, M. Liu, L. Mao, B. Kang, M. Xu, Y. Yu, S. Ermon, and W. Zhang, “MADiff: Offline multi-agent learning with diffusion models,” in Advances in Neural Information Processing Systems, 2024.arxiv.org/abs/2305.17330
- [8]D. Lee, D. Lee, and A. Zhang, “Multi-agent coordination via flow matching,” in International Conference on Learning Representations, 2026.arxiv.org/abs/2511.05005
- [9]P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,” in Advances in Neural Information Processing Systems, vol. 34, 2021.
- [10]Y. Shaoul, I. Mishani, S. Vats, J. Li, and M. Likhachev, “Multi-robot motion planning with diffusion models,” in International Conference on Learning Representations, 2025.arxiv.org/abs/2410.03072
- [11]J. Liang, J. K. Christopher, S. Koenig, and F. Fioretto, “Simultaneous multi-robot motion planning with projected diffusion models,” arXiv preprint arXiv:2502.03607, 2025.arxiv.org/abs/2502.03607
- [12]A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,” IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3861–3876, 2017.DOI
- [13]P. Glotfelter, J. Cortés, and M. Egerstedt, “A nonsmooth approach to controller synthesis for Boolean specifications,” IEEE Transactions on Automatic Control, vol. 66, no. 11, pp. 5160–5174, 2020.DOI
- [14]R. Lin and M. Egerstedt, ““Hierarchy of needs” for robots: Control synthesis for compositions of hierarchical, complex objectives,” in 2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 7682–7688.DOI
- [15]W. Xiao, T.-H. Wang, C. Gan, and D. Rus, “SafeDiffuser: Safe planning with diffusion probabilistic models,” in International Conference on Learning Representations, 2025.
- [16]D. Gadginmath, A. Allibhoy, and F. PasqualeLtti, “Provably safe generative sampling with constricting barrier functions,” Transactions on Machine Learning Research, 2026.arxiv.org/abs/2602.21429
- [17]X. Dai, Z. Yang, D. Yu, F. Liu, H. Sadeghian, S. Haddadin, and S. Hirche, “SafeFlow: Safe robot motion planning with flow matching via control barrier functions,” arXiv preprint arXiv:2504.08661, 2025.arxiv.org/abs/2504.08661
- [18]J. Yang, S. Jang, and S. Han, “SafeFlowMatcher: Safe and fast planning using flow matching with control barrier functions,” in International Conference on Learning Representations, 2026.arxiv.org/abs/2509.24243
- [19]R. Lin, M. Egerstedt, and F. PasqualeLtti, “Multi-agent flow matching with decoupled generative guidance,” arXiv preprint arXiv:2609.38133, 2026.arxiv.org/abs/2609.38133
- [20]J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems, vol. 33, 2020.arxiv.org/abs/2006.11239
- [21]Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in International Conference on Learning Representations, 2021.
- [22]Y. Du, C. Durkan, R. Strudel, J. B. Tenenbaum, S. Dieleman, R. Fergus, J. Sohl-Dickstein, A. Doucet, and W. Grathwohl, “Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and MCMC,” in International Conference on Machine Learning, ser. PMLR, vol. 202, 2023, pp. 8489–8510.arxiv.org/abs/2302.11552
- [23]M. F. Reis, A. P. Aguiar, and P. Tabuada, “Control barrier function-based quadratic programs introduce undesirable asymptotically stable equilibria,” IEEE Control Systems Letters, vol. 5, no. 2, pp. 731–736, 2021.DOI
- [24]S. Wilson, P. Glotfelter, L. Wang, S. Mayya, G. Notomista, M. Mote, and M. Egerstedt, “The Robotarium: Globally impactful opportunities, challenges, and lessons learned in remote-access, distributed control of multirobot systems,” IEEE Control Systems Magazine, vol. 40, no. 1, pp. 26–44, 2020.DOI
- [25]S. Zampini, J. K. Christopher, L. Oneto, D. Anguita, and F. Fioretto, “Training-free constrained generation with stable diffusion models,” arXiv preprint arXiv:2502.05625, 2025.arxiv.org/abs/2502.05625
- [26]A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola, “A kernel two-sample test,” Journal of Machine Learning Research, vol. 13, no. 25, pp. 723–773, 2012.DOI




