A robot that walks through a crowd has to be able to stop or step aside before anyone it cannot yet see steps into its path. How much room it needs for that depends on its body. A wheeled base can brake almost instantly. A quadruped carries momentum through a gait cycle. A humanoid must keep its center of mass over its feet while it decelerates. Today's safety filters ignore this difference. A safety filter here means a small optimization, solved many times per second, that minimally edits the planner's command so that a "keep-away" function stays positive; such functions are called control barrier functions (CBFs). These filters reason about a simplified point-mass robot and only about the pedestrians the camera currently sees. The cost is two familiar failures. On slow-to-stop bodies, the optimization has no valid answer at the moment a hidden pedestrian appears, so the solver fails or the robot collides. On agile bodies, uniformly inflated margins make the robot freeze.
We propose to measure, for each robot body, a conservative stopping envelope: how far and how long that specific machine travels while running an emergency stop or sidestep from its current speed and gait state. We then write this envelope directly into the barrier function, and enforce it against both tracked pedestrians and hypothetical "phantom" pedestrians placed at the edges of what the sensor cannot see. In plain terms, the robot only moves as fast as its own body can still stop before reaching any place a person might be, including places it cannot observe. The claim is that this one body-specific quantity makes the safety optimization feasible at every future step, not just the current one, without the blanket conservatism that causes freezing.
Core Claim
Optimization infeasibility in crowd navigation is mostly a property of the robot body interacting with partial observability, not of the crowd alone. If the barrier function subtracts a measured, body-specific stopping envelope from every clearance, then one simple fallback stays available at all times: brake or sidestep along the measured envelope. That fallback guarantees the filter always has a solution. The guarantee covers the observed pedestrians and any pedestrian that could emerge from an occluded region at a bounded walking speed.
Because humans are not controlled, the guarantee is deliberately limited to passive safety: the robot is at rest, or moving away, whenever contact could occur. We will not claim more than this.
Formally, for pedestrian or phantom $i$ at position $p_i$ and robot position $p$, we enforce
$$h_i(x) = \lVert p - p_i \rVert - r - d_e(v, \phi) - v_{h}\, t_e(v, \phi) \ge 0,$$
where $r$ is the combined body radius, and $d_e$ and $t_e$ are the distance and duration of embodiment $e$'s backup maneuver from forward speed $v$ and gait phase $\phi$. The term $v_h$ is an assumed maximum human walking speed. Personal-space comfort enters as a separate, softer barrier with a slack variable. That barrier is relaxed first, before any physical constraint, mirroring the priority-relaxation idea of Rule-based Optimal Control for Autonomous Driving, which drops low-priority driving rules when a combined quadratic program becomes infeasible. Unlike that work, our relaxation order is fixed and simple. The new element is the body-specific envelope and the occlusion phantoms, not the rulebook.
Where This Sits
Several papers motivate the gap. CE-Nav shows that a shared geometric velocity prior plus a short reinforcement-learning refinement per body transfers navigation across Go2, Spot, H1 and a quadrotor. It offers no safety certificate and tests only static obstacles. We reuse its two-stage planner as our nominal command source and add a certified layer below it. ASFM runs social navigation on Spot with onboard RGB-D and attributes its 15% failures in an obstructed hallway to intermittent detection dropouts. That is direct evidence that partial observability, not planning, is the residual failure. The Focused Risk Perception work reports that its TurtleBot policy degraded on hardware because motors could not match simulated agility, which is the embodiment gap showing up in crowds.
Other papers supply methods we build on or compete with. The optimal-decay CBF of Safety-Critical Control using Optimal-decay Control Barrier Function with Guaranteed Point-wise Feasibility guarantees a solution at each instant by letting a slack term soften the barrier's decay rate. The authors themselves note this is point-wise only and can lead into dead-ends. Our envelope targets exactly that persistence gap. Online Control Barrier Functions for Decentralized Multi-Agent Navigation and Learning Environment-Aware Control Barrier Functions learn a graph network that retunes barrier parameters online with a feasibility reward. They assume homogeneous cooperative robots with full neighbor states inside a sensing radius. We assume heterogeneous bodies, uncooperative humans and occlusion. Pointwise Feasibility of Gaussian Process-based Safety-Critical Control handles model uncertainty by enlarging margins. Its feasibility conditions explain why uniform inflation can swallow the admissible control set, and that conservatism is the failure we expect our body-specific envelope to avoid.
Safe Diffusion MPC forecasts joint crowd futures and prunes unsafe samples with reciprocal velocity obstacles. It tests wheeled robots in crowds of two to three people and offers no feasibility proof under actuator saturation. It is complementary, since its forecasts could tighten $v_h$ for observed people. We do not depend on it.
Research Program
Envelope measurement. The first component is a calibration procedure. For each body (Go2, Spot, H1 in Isaac Sim, plus a differential-drive base), we sample thousands of initial states over speed, yaw rate and gait phase. At each state we trigger the body's native emergency-stop and lateral-sidestep behaviors and record the swept footprint and stop time. Using the platforms' existing locomotion controllers keeps the measurement honest about actuator lag and balance limits. We then fit $d_e$ and $t_e$ as a high-quantile regression on these rollouts (99th percentile, plus a validation-set check that coverage holds). We expect this to cost compute on the order of the CE-Nav refinement budget, but we will report the actual cost. Domain randomization over mass and latency gives the envelope some margin for sim-to-real mismatch.
Observability handling. The second component handles what the robot cannot see. A tracker keeps each person whose detection drops out, with a position uncertainty disc that grows at $v_h$ per second. This directly targets the ASFM dropout failure. Within sensor range, every occlusion boundary (the edge of the field of view, and the shadows of walls and of people) spawns phantom pedestrians on the boundary, moving toward the robot at $v_h$. To avoid an explosion in constraint count, only the eight most critical constraints by time to collision enter the optimization, following the top-$K$ result of the Focused Risk Perception paper.
Filter. The third component is the filter itself: a quadratic program over the body-frame velocity command, with hard barriers from the equation above, soft proxemic barriers, and per-body command limits. We will prove that the backup maneuver satisfies all hard constraints whenever they held at the previous step, which gives persistent feasibility under the passive-safety notion. We will also state plainly the assumption this rests on: humans never exceed $v_h$.
Baselines
The naive reference is CE-Nav's planner and refiner run directly on each body with no safety filter. Its reported mean success of 0.8575 on Go2 and 74.5% on H1 in obstacle forests (CE-Nav) establishes that the nominal planner is competent. Any collisions it then produces in our crowd scenes come from the missing filter, not a weak planner. It is run with the published two-stage training at the same per-body budget as our calibration. We beat it if we cut collisions with emerging pedestrians while keeping success within a few points.
The strongest existing method on the feasibility question is the learned online-CBF parameter tuner of Online Control Barrier Functions for Decentralized Multi-Agent Navigation. It is the only cited approach that learns to keep the safety program feasible in bottlenecks, and it reportedly solved narrow passages that an exhaustive grid search over fixed parameters could not. No scalar success value is reported in the file, so there is no reported result on Lacuna for its success rate on these benchmarks. For fairness, we retrain it on our crowd scenes with the same simulation steps we spend on calibration. We give it the tracked pedestrians as neighbors, and in a strengthened variant we also give it a body identifier and the phantom agents. Beating the strengthened variant is what would show the envelope itself, rather than our observability handling, is the active ingredient.
We also run the optimal-decay CBF with a point-mass model, and a uniformly inflated robust margin tuned per body by grid search to zero collisions. These are the "just soften it" and "just pad it" references.
Three ablations isolate our ingredients. The first replaces the measured envelope with an analytic point-mass braking distance, which isolates the body-specific measurement. The second removes the phantoms, which isolates the occlusion handling. The third uses one shared envelope across bodies, which tests whether per-body calibration matters at all.
Why not just…?
Why not just add a collision penalty to the CE-Nav refiner and let it learn caution per body? This is a real threat on average-case metrics, because the refiner already adapts to momentum and latency. Evidence from SocNavBench shows an end-to-end learned policy (SA-CADRL) had the most collisions, 40, against 64 for a pedestrian-unaware planner. That suggests learned caution is unreliable in rare encounters. We will settle the question by training this variant at matched budget and reporting tail statistics: collisions with pedestrians who emerge from occlusion at walking speed, a situation that is rare in training.
Why not just inflate the safety margin and slow down near corners? A uniformly tuned margin can reach zero collisions, but it must be set for the slowest-stopping body or tuned per body. Uniform inflation is exactly the conservatism the Gaussian-process feasibility analysis warns can empty the admissible control set. Our grid-searched inflation baseline will show whether it produces more freezing time and longer time-to-goal than the envelope at equal collision rates. If it does not, the envelope adds nothing and we will say so.
Why not just use slack variables or optimal-decay CBFs so the program is always solvable? Slack makes the program solvable by weakening the guarantee at the moment it matters. The optimal-decay paper itself stops at point-wise feasibility. Our comparison will report how often slack is actually consumed on hard barriers, and whether those steps precede collisions.
Proposed Experiments
We build an Isaac Sim crowd suite on the CE-Nav obstacle-forest scenes. It populates the scenes with reactive social-force pedestrians seeded from ETH/UCY and ATC mall trajectories, and adds doorways, corners and the four Gazebo patterns (crossing, head-on, following, random) from the Focused Risk Perception benchmark. We use reactive pedestrians because the SocNavBench authors show that non-reactive replays create unavoidable collisions. We additionally run SocNavBench's 33 episodes on the wheeled body for comparability. The reported reference points are Social Forces at 32/33 successes and SA-CADRL with 40 collisions. On Gazebo head-on, the TD3 top-$K$ method reports 93.3% success.
Metrics are taken per body:
- fraction of control steps where the program is infeasible or uses hard slack;
- collisions split into observed versus emerged-from-occlusion pedestrians;
- passive-safety violations;
- freezing time;
- success and time-to-goal;
- personal-space intrusion duration.
No reported result was found on Lacuna for any method on occlusion-emergence scenarios, so these will be new measurements.
The decisive checks are three. First, on H1 and Spot, the measured-envelope filter should show near-zero infeasible steps and emergence collisions, while the point-mass-envelope ablation should not. If both behave the same, the body-specific measurement is unnecessary. Second, at equal collision rate, the envelope should freeze less than uniform inflation on Go2 and the wheeled base. Third, a calibration-coverage test on held-out rollouts with perturbed mass and latency should show whether the 99th-percentile envelope still holds. This is where sim-to-real risk would surface before any hardware trial. A small hardware study on Go2 in a hallway with a blind corner, mirroring the ASFM setup, would follow only if all three checks pass.
References
- Rule-based optimal control for autonomous driving
- CE-Nav: Flow-Guided Reinforcement Refinement for Cross-Embodiment Local Navigation
- ASFM: Augmented Social Force Model for Legged Robot Social Navigation
- Crowd Navigation for Mobile Robots with Focused Risk Perception
- Safety-Critical Control using Optimal-decay Control Barrier Function with Guaranteed Point-wise Feasibility
- Online Control Barrier Functions for Decentralized Multi-Agent Navigation
- Pointwise Feasibility of Gaussian Process-based Safety-Critical Control under Model Uncertainty
- Safe Diffusion Model Predictive Control for Interactive Robotic Crowd Navigation
- SocNavBench: A Grounded Simulation Testing Framework for Evaluating Social Navigation