No-Fault Adversaries: Training Crowd Navigation Against Hostile but Avoidable Pedestrians

Robots that move among people are mostly trained in simulators where the other agents are polite. In those simulators the pedestrians follow fixed avoidance rules or scripted paths. In the real world, people cut corners, stop abruptly, and step out from behind each other. A natural fix is adversarial training: let a learned opponent control the pedestrians and push the robot toward collisions. This fails immediately, because a pedestrian that wants to hit the robot always can. The robot then learns to stay still, which is the well-known "freezing robot" problem. We propose to train the pedestrian adversary under a no-fault rule. At every step, the adversary may only choose motions that leave the robot at least one escape maneuver, such as braking or swerving, that stays collision-free. Every collision the adversary produces is therefore one the robot could have avoided. The adversary becomes a hunter of genuine policy mistakes rather than an unbeatable rammer.

We pair this adversary with randomization of the robot's own dynamics and perception: latency, acceleration limits, and noise and dropouts in the detected positions of people. The goal is a policy that transfers from simulation to a physical robot without fine-tuning. The bet is that this targeted, constrained adversary finds the rare timed cut-ins and occlusion ambushes that uniform randomization almost never samples. It should do so without the conservatism that worst-case training is known to cause.

Core Claim

Three bodies of work each stop short of this combination. Adversarial training (Robust Adversarial Reinforcement Learning, RARL) applies learned destabilizing forces to the robot's own links. It works only because those forces are hand-bounded, and its authors note that an unconstrained adversary prevents any learning. Minimax-regret environment design (Replay-Guided Adversarial Environment Design) chooses training levels where the student's gap to an optimal policy is largest. Its levels are static layouts: maze walls and fixed track shapes. Refining Minimax Regret for Unsupervised Environment Design shows such adversaries stall on levels whose regret cannot be reduced. Crowd navigation methods reject the static-world assumption but train against cooperative, rule-following humans. Examples are Crowd-Robot Interaction (LM-SARL), Hyp2Nav, and SafeCrowdNav.

Our claim is that a simple, checkable rule can stand in for regret estimation when the adversary controls moving obstacles. The rule is an "inevitable collision state" test: is the robot in a state from which no available maneuver avoids a crash? Ramming then becomes impossible by construction. Adversarial pressure then improves both safety and progress on held-out hostile crowds, without slowing the robot in ordinary crowds.

The rule borrows its logic from static-world safety work. Safe Motion Planning in Unknown Environments compares a robot to a safe agent that never enters an inevitable collision state with respect to unseen obstacles, and uses emergency braking to guarantee it. Evaluating Trajectory Collision Probability through Adaptive Importance Sampling shows that safety certificates computed on static geometry fail once things move. Both papers list dynamic obstacles as future work. We turn their notion of inevitable collision into a constraint on the opponent rather than a guarantee for the planner.

The Adversary's Game

Let $s_t$ be the joint state of robot and pedestrians at step $t$. Let $\mathcal{E}$ be a small fixed library of robot escape maneuvers: full brake, and hard swerves left and right under the randomized acceleration limits. Each adversarial pedestrian commits to its next $H$ steps of motion. The adversary's action $u_t$ is admissible only if at least one maneuver in $\mathcal{E}$ stays collision-free against that committed motion. Each pedestrian must also keep a minimum speed toward its own sampled goal, so it cannot win by standing in a doorway.

The adversary maximizes

$$ R_{\text{adv}} = \mathbb{1}[\text{collision}] + \lambda \, \frac{T_{\text{robot}} - T_{\text{ref}}}{T_{\text{ref}}}. $$

Here $T_{\text{robot}}$ is the robot's time to goal, with timeouts counted as the episode limit. $T_{\text{ref}}$ is the time on the same start and goal with a non-adversarial crowd. $\lambda$ trades off the two terms. Rewarding delay matters: freezing becomes a failure the adversary can exploit, so the robot cannot buy safety by becoming timid.

In simulation we have privileged access to the full state, so admissibility is a cheap forward rollout of $|\mathcal{E}|$ maneuvers. We implement it as a projection: inadmissible adversary actions are replaced by the nearest admissible one. As in RARL, robot and adversary are updated in alternation. A mixed-population schedule keeps a fraction of pedestrians on ORCA, the standard rule-based reciprocal avoidance model, so the robot still sees ordinary crowds.

Research Program

We build on the CrowdNav simulator and the hyp2nav codebase, which contains attention-based and hyperbolic crowd policies. This lets us hold the robot architecture fixed and vary only the training distribution.

The work proceeds in three stages.

The first stage implements the no-fault adversary and verifies three things. Admissibility must be cheap enough not to dominate training compute. Projected actions must still produce diverse behaviors, including cut-ins, sudden stops, and occlusion of one pedestrian by another. Collision reward under the constraint must be learnable, unlike the unconstrained case.

The second stage adds dynamics and perception randomization for the robot. We randomize command latency, acceleration and velocity limits, wheel slip, and detection noise, dropout, and false positives on pedestrian states. We also let the adversary choose when to occlude itself from the robot's simulated sensor. Perception corruption then becomes a structured attack rather than only uniform noise. This responds to the gap identified in S2R-Bench: synthetic corruptions underestimated real glare effects on LiDAR by 85%, and sensor misalignment caused an 81–82% drop. Uniform noise misses structured shifts.

The third stage deploys the policy zero-shot on a physical differential-drive robot. Pedestrian states come from an onboard detector, following the hardware setting of Safe Diffusion Model Predictive Control, which used 2–3 moving humans.

We also analyze the "safety tax." On the Regret of H∞ control proves that worst-case controllers pay a performance cost in benign conditions, and that this cost grows with the distance between the real disturbance and the assumed worst case. Our constraint shrinks the assumed worst case to the avoidable set. We therefore predict a smaller nominal slowdown than unconstrained or simply bounded adversarial training. We test this by measuring navigation time on standard ORCA crowds.

Baselines

Naive reference. The do-nothing reference is the same robot architecture trained on standard CrowdNav ORCA crowds with no adversary and no randomization. It shows what the default training distribution delivers.

Strongest existing method. This is Hyp2Nav, retrained with its released code. On CrowdNav, Hyp2Nav reports 99.8% success in Square scenarios, compared with SafeCrowdNav's 97.7%, and 94.6% success in a 20-human stress test. SafeCrowdNav reports 98.6% success and 9.984 s navigation time in circular scenes. LM-SARL reports 100% success in the setting where humans ignore the robot, and 10.59 s navigation time in the setting where they react to it, versus 12.29 s for ORCA. These near-ceiling numbers are exactly why we need harder held-out tests. Beating the strongest baseline means matching its nominal success within one point, and keeping navigation time within 5%, while substantially reducing collisions on the held-out hostile suite and on hardware. We found no reported result on Lacuna for any of these methods under adversarial pedestrians or on physical hardware.

Safety-filter baseline. We add the moving-obstacle safety filter of Safe Reinforcement Learning of Robot Trajectories in the Presence of Moving Obstacles. It combines a backup avoidance policy, a learned risk network, and limits on robot jerk, and it extended collision-free time from 2.7 s to over 40,000 s in deterministic simulation against stochastic obstacles. We apply it on top of the naive policy.

Plain RARL. This is RARL transposed to pedestrians. It gives the adversary the same collision-plus-delay reward with only a pedestrian speed bound, which matches RARL's use of a hand-tuned force budget.

Ablations. The key ablation keeps the admissible action set but replaces the learned adversary with uniform sampling inside it. This separates the value of targeting from the value of the constraint.

Fairness and success criteria. All methods get the same environment-step budget, the same robot network, the same randomization of robot dynamics and perception, and the same hyperparameter search of eight configurations, with five seeds each. Success means statistically separating from all baselines on the held-out hostile suite while staying within the nominal margins above.

Why not just…?

Why not just randomize pedestrian behavior uniformly? We could sample ORCA parameters, non-reactive walkers, random stops, and aggressive speeds, as standard randomization does for physics. This is the most serious threat. RARL's evidence that learned adversaries beat random perturbation comes from force disturbances on the robot's own body, not from crowds. Likewise, the result in Replay-Guided Adversarial Environment Design, where curated levels beat random ones using under 0.25% of environment steps, concerns static track layouts. Our "uniform-within-admissible-set" ablation settles the question directly. If it matches the learned adversary on the held-out hostile suite, the contribution reduces to the constraint itself, and we will report it as such.

Why not just use minimax-regret level curation over crowd scenario parameters? We could run PLR⊥ or ReMiDi on spawn positions and behavior parameters instead of training a per-step adversary. That treats a crowd as a static level, chosen once per episode, and cannot react to the robot within an episode. Timed cut-ins require that kind of within-episode reaction. ReMiDi also shows that regret estimates fail when some levels contain unavoidable failures. Ramming crowds are exactly such levels, and our rule removes them without estimating regret. We include PLR⊥ over scenario parameters as an extra comparison so that this point is tested rather than asserted.

Why not just add a safety filter to a nominally trained policy? The filter switches between the task policy and a backup policy. Its authors note this produces conservative or jerky motion, and it assumes perfect obstacle states. A filter can only veto actions after the fact. It cannot teach the task policy to stay out of situations where the filter must keep intervening. That is what our delay term rewards the adversary for exposing. We measure filter intervention rates and timeouts; if the filtered policy matches us on both, the adversary is unnecessary.

Why not just use a modular planner or ORCA? Waypoint modularity transferred road-following zero-shot: Driving Policy Transfer via Modularity and Abstraction reports 100% versus 0% for end-to-end control. However, its authors state that it does not handle dynamic agents. ORCA is slower than learned policies (12.29 s vs 10.59 s in LM-SARL). We keep ORCA as an evaluation reference rather than a main competitor.

Proposed Experiments and Decisive Checks

Evaluation suites. We evaluate on three simulation suites. The first is standard CrowdNav circle- and square-crossing with 5, 10, and 20 humans, measuring success, collision rate, timeout rate, navigation time, and personal-space intrusions. The second is a held-out hostile suite generated by an independently trained no-fault adversary, using a different seed and a different $\lambda$. No method trains against it. Every collision in this suite was avoidable by construction, so it measures genuine policy mistakes. The third replays ETH/UCY pedestrian trajectories, which are recorded real human walking, under the same perception corruption.

Hardware. We run a fixed protocol of natural walking plus scripted cut-ins and occluded emergences. Following the setting of Safe Diffusion Model Predictive Control, we use 2–3 humans. We report collisions, emergency stops, and time to goal, with the same trial count for every method.

Decisive checks.

The first check is the core claim: the learned no-fault adversary must beat both uniform sampling inside the admissible set and plain RARL on the hostile suite, without losing more than one point of success on standard CrowdNav.

The second check is freezing: plain RARL should show sharply higher timeout rates, and the no-fault variant should not.

The third check is the safety tax: on ORCA crowds, navigation time for our policy should stay within 5% of the naive policy, while plain RARL's should not.

The fourth check is transfer: gains on the hostile suite should carry over to hardware cut-in trials. If they vanish, the bottleneck is perception mismatch rather than behavior, which would redirect the work toward the structured sensor failures documented in S2R-Bench.