Who Dodges Whom When Robots Don't Know Their Own Limits

When an agile robot and a heavy robot are about to collide, the sensible rule is that the agile one should do most of the dodging. A well-known controller for mixed robot teams does exactly this. Each robot runs a small optimization that minimally edits its commanded motion so that a "barrier" condition on inter-robot distance keeps holding. The obligation to keep that condition is split between the two robots in proportion to how hard each can brake or accelerate. The split only works if those capabilities are known numbers. Real robots lack fixed limits. A rover on a polished floor, carrying a payload, with a sagging battery and a 100 ms actuation lag, can stop far less sharply than its datasheet says. When a robot's claimed capability is too high, it accepts more of the dodging than it can physically perform, and the safety argument fails exactly when it is needed.

We propose to train, in heavily randomized simulation, a small recurrent estimator that watches each robot's own recent commands and motion. From that history it outputs a calibrated, conservative bound on what the robot can currently do: a lower bound on achievable deceleration and an upper bound on actuation delay. A held-out set of simulated physics settings calibrates the bound so that it covers the true value with a chosen probability. Robots broadcast these bounds. The pairwise safety filter uses them both to split avoidance responsibility and to size safety margins. Moving non-cooperative obstacles such as pedestrians or uncontrolled vehicles are assigned zero responsibility, so the team must absorb the full avoidance burden. A learned multi-agent policy then explores freely underneath this filter. The bet is simple. Domain randomization, which is usually used to make a policy robust, is better spent making the safety filter's beliefs about the robot itself accurate. That should give near-worst-case safety without worst-case sluggishness.

Core Claim

The weak point of asymmetric multi-robot safety filters under the reality gap is the capability numbers that decide who yields, more than the filter itself. Safety Barrier Certificates for Collisions-Free Multirobot Systems supplies the method we build on. It introduces "Strategy C," which splits each pairwise collision-avoidance constraint inversely to agility. It also proves the safe set stays invariant, and it estimates neighbors' acceleration bounds online by least squares from conservative initial values. It assumes, however, a static surrounding environment, delay-free unicycle dynamics, and that each robot's own limits are known constants. Watching a neighbor accelerate shows what it has done. It does not show what it can do now on a slippery floor, which is the quantity safety actually depends on.

We claim three things. First, a capability estimator trained across randomized friction, mass, payload, battery and latency can produce bounds that are both conservative and tight. Second, plugging these bounds into Strategy C, together with a delay-aware margin and full-responsibility handling of moving obstacles, keeps collisions near the level of a filter using worst-case constants. Third, it recovers most of the task speed that worst-case constants throw away.

Concretely, for robots $i$ and $j$ with barrier function $h_{ij}(x)$ (positive when the pair can still brake to a stop without touching), Strategy C requires each robot to satisfy its share of the barrier condition:

$$ \text{robot } i \text{ enforces: } \; \dot h_{ij}(x,u_i) + \rho_{ij}\,\gamma\, h_{ij}(x) \ge 0, \qquad \rho_{ij} = \frac{\underline{a}_i}{\underline{a}_i + \underline{a}_j}. $$

Here $\gamma>0$ is the usual barrier gain. $\underline{a}_i$ is robot $i$'s calibrated lower bound on achievable deceleration, which replaces the nominal constant. The braking distance inside $h_{ij}$ is computed with $\underline{a}_i,\underline{a}_j$ and the calibrated delay bound $\bar\tau_i$, so that $\bar\tau_i$ inflates the margin. For a non-cooperative obstacle $o$ we set $\rho_{io}=1$ and propagate its reachable set forward over $\bar\tau_i$ plus the braking time.

That last choice targets the inevitable-collision-state problem raised by Schmerling et al. and Janson et al.. Both show that a state which is safe against a frozen snapshot of the world can already be doomed once moving obstacles are accounted for.

Research Program

Simulator. The first piece is a continuous, multi-robot simulator with genuinely mixed embodiments: light differential-drive robots, a heavy rover, and optionally a small quadrotor. Each episode randomizes per-robot floor friction, payload mass, motor saturation, battery sag and actuation delay. Each episode also contains scripted moving obstacles with randomized speeds and goal changes. Parameter ranges deliberately include the low-friction, high-delay regimes where nominal Strategy C should fail. We will design the simulator so that the Safety Gymnasium cost conventions (per-step collision cost, normalized cumulative cost) apply directly.

Estimator. The second piece is the capability estimator. A small recurrent network per robot embodiment reads the robot's last few seconds of commanded and measured velocity. It predicts deceleration capability and delay, with ground truth computed from the simulator's sampled parameters. Its output is turned into a one-sided bound using split-conformal calibration over held-out randomized domains. Calibration gives finite-sample coverage $1-\alpha$, but only for domains drawn from the same distribution as calibration. The real world does not satisfy that assumption.

Out-of-distribution guard. The third piece covers that gap. Following Safe Domain Randomization via Uncertainty-Aware Out-of-Distribution Detection and Policy Adaptation, which uses disagreement among a diversity-trained critic ensemble to flag unfamiliar dynamics on a single quadruped, we train a small ensemble of estimators. When their disagreement exceeds a threshold set in simulation, the robot falls back to worst-case constants. We reuse that paper's detection idea, but aim it at a safety filter's capability inputs rather than at a locomotion policy's value function.

Learning under the filter. The fourth piece is task learning beneath the filter. We train a cooperative policy with a constrained multi-agent learner of the kind in Scalable Constrained Policy Optimization for Safe Multi-agent Reinforcement Learning. The task is heterogeneous coverage-and-delivery, where the heavy rover carries payloads and the light robots scout. The filter corrects every action during both training and execution.

Analysis. We will measure coverage of the capability bound in held-out simulated domains and how often the fallback fires. We will also record each robot's share of the total filter correction, to check that responsibility actually moves toward whichever robot is currently agile. For example, a light robot on a slippery patch should start yielding less.

Hardware. Where a mixed-robot lab is available, we will run the system zero-shot on a setup like the Khepera III plus Magellan Pro testbed of the barrier-certificate paper. A person or a remote-controlled cart will serve as the moving obstacle, and we will add low-friction mats and payloads. Hardware access is the main practical risk, and the simulation results must stand on their own.

Relation to Prior Work

The closest bridging work motivates the gap without filling it. Model-based Dynamic Shielding for Safe and Efficient Multi-Agent Reinforcement Learning merges per-agent safety monitors when agents come close. It learns dynamics with a recurrent world model and keeps safety above 90% from 4 to 12 agents in particle environments. Its agents are kinematically alike, it never transfers out of simulation, and its own authors note that its guarantees fail when the learned model mispredicts. We learn only a scalar capability bound, which is much easier to calibrate than a full world model, and we inherit the barrier proof for everything else.

Probabilistic Safety Barrier Certificates handles bounded sensing and motion noise for homogeneous teams. It does not deal with uncertainty in the parameters that set responsibility.

Sim-to-Lab-to-Real and LyEvO give certified or shielded sim-to-real transfer for one robot among static obstacles. LyEvO pays for its certificates with more than 31 hours on 1,000 CPU cores. Provably Safe Sim-to-Real Transfer achieves 0% unsafe iterations by shrinking a "mismatch region." It does so only for tabular single-agent problems. All of these constrain our ambition, and none touches the question of who yields.

Two further works constrain the design. Shield Decentralization for Safe Multi-Agent Reinforcement Learning shows that decomposed, communication-free shields can be exactly safe. Its decomposition assumes synchronous, identical agents, which is why we handle delay inside each pairwise margin rather than through a joint factorization. Safety Gymnasium reports that a Lagrangian learner cut violations 98% at a 45% reward cost. That trade-off suggests a learner alone will not reach hardware-grade safety.

Baselines

Naive learned baseline. The do-nothing reference is a multi-agent policy trained with the same domain randomization and a collision penalty, with no filter. A Lagrangian-constrained variant (PPO-Lag style, as shipped in SafePO) is the stronger version of this reference. It shows what randomization plus a learned constraint achieves alone. The comparison is fair because it gets the same simulator, the same number of environment steps and the same tuning budget of about twenty configurations. For the Lagrangian variant the closest reported figure is the Safety Gymnasium result above, a 98% violation reduction with a 45% reward drop. That figure comes from single-agent velocity tasks, and there is no reported result on Lacuna for heterogeneous teams with randomized dynamics.

Strategy C with nominal constants. This is the barrier-certificate controller exactly as published, with datasheet limits and its online least-squares neighbor estimation, wrapped around the same learned task policy. The original paper reports reaching all goals "without a single collision" in 6-agent simulation. It has no reported result on Lacuna under randomized friction, payload or delay, which is precisely the regime we test.

Strategy C with worst-case constants. This variant takes the minimum deceleration and maximum delay over the whole randomization range. It is the strongest safety competitor and our real target. We expect it to match our collision rate. Beating it means equal or lower collisions and infeasibility with substantially shorter task completion time. We fix in advance that recovering at least half of the completion-time gap between worst-case and nominal constants counts as a win.

Symmetric 50/50 barrier. This uses the same calibrated bounds but splits responsibility equally. It isolates whether the asymmetric split still matters once margins are calibrated.

Model-based dynamic shielding. We reimplement MBDS on our environment, with its world model trained on the same randomized data. It is the strongest existing learned shield.

Ablations. Three ablations isolate the key ingredient. The first uses the estimator's mean prediction instead of the calibrated bound, to test whether calibration rather than estimation carries the result. The second removes the delay term. The third removes the out-of-distribution fallback.

Why not just…?

Why not just use worst-case limits over the randomization range? This is the most serious alternative. On collisions alone it may well tie us. Its cost is conservatism. Robots that are usually agile get treated as permanently sluggish, so they yield late, keep large gaps, and freeze more often in crowded crossings. The deadlock and infeasibility problems already reported for Strategy C should therefore get worse. The comparison of completion time and deadlock rate against this baseline settles the question. If the gap is small, for example because the randomization range is narrow in practice, then calibration is not worth its complexity and we will say so.

Why not just rely on Strategy C's own online estimation of neighbor limits? That estimator learns from observed motion, so it only certifies what a robot has already done. It cannot detect a sudden loss of traction or a newly added payload before the robot needs to brake. It also says nothing about the ego robot's own delay. The "Strategy C with nominal constants" baseline includes this estimator, and the held-out low-friction domains test exactly this failure.

Why not just train a domain-randomized policy with a big collision penalty? Quanyi Li et al. document the two failure modes of penalty tuning: penalties that are too small accept crashes, and penalties that are too large produce robots that stand still. MBDS reports penalty-based methods losing 10-20% safety as agent count grows. In Shield Decentralization, unshielded agents with momentum averaged over 115 violations per 100 episodes. Our naive baseline measures this directly on heterogeneous teams.

Why not just treat moving obstacles as static and replan fast? A snapshot of the obstacle cannot account for the robot's own braking distance plus actuation delay, so it walks into inevitable-collision states. We run this variant, with the snapshot treatment and $\rho_{io}$ computed from current position only, as an extra ablation in the moving-obstacle scenarios.

Proposed Experiments

Simulation. All methods train on the same randomized distribution. We evaluate three settings, each for 4, 8 and 12 robots in dense crossing and delivery layouts with zero to four moving obstacles: in-distribution domains, held-out domains drawn from the same distribution, and shifted domains with friction and delay beyond the training range. The metrics are collisions per hour of operation, the fraction of steps where the filter's optimization has no solution, the deadlock rate, task completion time and reward, the per-robot share of filter correction, the empirical coverage of the capability bound, and the fallback rate.

Decisive checks. Three outcomes would decide the proposal. On held-out low-friction domains, nominal Strategy C should collide while our method should not. If both are safe, the premise is wrong. On the same domains, our method should approach the worst-case baseline's collision rate while finishing markedly faster. If it does not, calibration buys nothing. On shifted domains, the out-of-distribution guard should fire before collisions rise. If coverage collapses silently, the method is unsafe beyond the training range, and the paper must report that limitation rather than claim transfer.

Hardware. Any physical runs reuse the frozen estimator and filter with no real-world fine-tuning. They report the same metrics over repeated crossings with payload and low-friction changes and a human-driven moving obstacle.