Learning Side-by-Side Navigation for Personal Mobility Vehicles from Human Drivers

Luis Yoichi Morales; Florent Ferreri

cs.ROOct 27, 2026

Learning Side-by-Side Navigation for Personal Mobility Vehicles from Human Drivers

Authors: Luis Yoichi Morales, Florent Ferreri

Abstract

This paper presents an approach to achieve side-by-side autonomous motion with a pair of personal mobility vehicles (PMVs). Given a global path, two PMVs plan cooperatively so as to hold a social formation that lets the passengers hold a conversation, while keeping the ride safe and smooth. Each vehicle runs its own sampling-based model-predictive planner; the planners exchange nothing but the odometry each vehicle could observe of the other, and the joint decision is expressed as a pair of utility matrices over pairs of candidate trajectories, resolved by a common rule. The behavioural parameters are fitted by inverse reinforcement learning from 409 minutes of driving by ten pairs of people in human-driven PMVs in an indoor environment, and the lanes the pair can use are derived offline from the corridor itself rather than fixed by hand. A safety layer the cost cannot override executes a move only if the vehicle can still brake clear, measured from its whole body, and a vehicle that is stuck reverses along its own path and continues. In closed-loop simulation over 24 runs in the conditions the drivers were recorded in and 24 in four generated obstacle layouts, no vehicle entered a wall or a cone. The pair holds conversational formation for 60.5 and 68.9 percent of the time among obstacles (anticlockwise and clockwise) and 81.2 and 91.5 percent in a clear aisle, against the drivers' 34.0 and 36.0 percent, and 52.9 and 56.0 percent, and it matches their speed. It keeps one lateral formation where the drivers alternated between a close diagonal and a wide abreast one, and it varies far less. Experimental results are based on a differential drive kinematic simulator running on ROS 2.

Figures & tables

Figure 1: Two personal mobility vehicles performing shared planning to move side by side enabling the passengers to engage in a conversation.

Figure 2: PMV’s system comfiguration. Each vehicle is equiped with a $3D$ LiDAR and wheel encoders. Vehicle parameters are shared through wifi.

Figure 3: Trajectories of PMV’s moving CW while the drivers were engaged in a conversation in an indoor environment. The color represents the velocity of each vehicle. Some cones were placed as obstacle in half of the experimental conditions.

	anticlockwise		clockwise		total
condition	runs	min	runs	min	runs
clear aisle	21	99	21	102	42
surveyed cones	21	100	22	108	43
total	42	198	43	211	85

Table I: The human-driven corpus: recordings and minutes of paired driving per condition and direction,

10

pairs.

term	measured as	$\sigma$	weight
bearing $R_{a}$ (signed)	$\langle\|\sin R_{a}-s\|\rangle$	0.133	0.53
separation (social distance)	$\|d_{T}-d^{\star}_{R}(o)\|$	0.3 m ‡	2.00
speed	$\langle\|v-v^{\star}(\kappa)\|\rangle$	0.2 m/s ‡	3.00
acceleration	$\langle\|\dot{v}\|\rangle$	0.2 m/s 2	1.05
path distance	$\langle\mathrm{dist}(\cdot,\mathrm{lane})\rangle$	0.3 m ‡	1.91
cross-track (signed)	$\|\langle c\rangle-c^{\star}\|$	0.203 m	6.00

Table II: The utility terms the planner uses.

Figure 4: What the drivers did: one panel per term of Table II , the raw quantity pooled over the ten recorded pairs, with the corpus median in orange and the target the cost asks for dashed — a shaded band where that target is a function of context rather than a constant. For the signed cross-track position the target is the vehicle’s lane offset, so the two bands are the two lanes; progress is the only reward and has no target.

R_{a}

follows the definition of our earlier side-by-side work,

R_{a}=\mathrm{ang}(p_{i})-\mathrm{ang}_{i}(p_{j},p_{i})

, and peaks just under

90^{\circ}

. Two distributions carry their own warnings. motion_speed has a hard spike at

1.24

m/s, which is the vehicle’s governor and not a preference, so the drivers’ speed is censored from above. rel_dist_end is bimodal , at

0.85

and

1.50

m, because the pair switches between riding abreast and filing, and the two formations hold different distances; the split is within pairs, not between them. Fitting a unimodal target to it averages across two behaviours and lands between them.

	surveyed cones				clear aisle
	ccw		cw		ccw		cw
% of time	drv.	ours	drv.	ours	drv.	ours	drv.	ours
side by side	34.0	60.5	36.0	68.9	52.9	81.2	56.0	91.5
filing	48.0	37.3	45.8	28.4	37.1	18.8	36.1	7.3
apart $\geq 2$ m	17.2	2.1	17.8	1.3	9.2	0.0	7.0	0.3
one stopped or reversing	0.8	0.1	0.4	1.5	0.8	0.0	0.9	0.9

Table III: Share of time by what the pair was doing, drivers (drv.) against the planner (ours), in the two conditions the drivers were recorded in. Side by side: the partner within

25^{\circ}

of abreast and under

2

m away.

clear aisle
		simulator / drivers		Wasserstein distance
dir	quantity	mean	s.d.	sim vs. drivers	driver vs. drivers
ccw	lateral sep.	1.17 / 1.09	0.14 / 0.45	0.28	0.08
ccw	along-track	0.35 / 0.59	0.29 / 0.55	0.24	0.13
ccw	speed	0.87 / 0.91	0.24 / 0.25	0.05	0.11
cw	lateral sep.	1.15 / 1.01	0.18 / 0.44	0.27	0.09

Table IV: Wasserstein distance from our distribution to the pooled drivers, against the spread among the drivers themselves: the median distance from one recorded pair to the pool of all the others (

10

pairs per direction and condition).

Figure 5: Clear aisle. Pooled driver behaviour filled, each recorded pair a thin line, ours over the top. A curve inside the spread of the thin lines is not distinguishable from a human pair on that quantity.

Figure 6: Surveyed cones, same construction.

condition	dir	runs	wall contact		cone contact		stuck	backups	holds	stopped
			runs	worst	runs	worst	runs	/run		s/run
				mm		mm
surveyed cones	ccw	6	0	0	0	0	0	0.2	1	0
	cw	6	0	0	0	0	0	2.2	0	6
clear aisle	ccw	6	0	0	—	—	0	0.0	0	0
	cw	6	0	0	—	—	0	0.2	1	2

Table V: Contact, liveness and recovery. Contact is measured from the logged poses, independently of the planner, against every surveyed map cell and every placed cone; a run counts when the body entered a wall or a cone by more than

25

mm, and worst is the deepest penetration. Stuck: runs ending with a vehicle stopped.

concept	mean	median	vs. oracle
oracle (action-space floor)	1.058	0.256	—
optimistic	1.603	1.120	$+0.544$
social	1.645	1.145	$+0.587$
maximin	1.663	1.159	$+0.605$
anticipated	1.757	1.307	$+0.699$
random (chance)	2.014	1.705	$+0.956$

Table VI: Open-loop agreement with the drivers: distance between the chosen rollout’s endpoint and the drivers’ realised position 3 s later, over 300 recorded timesteps.

vs. social	cone contact	stalls/run	progress
optimistic	$+4.5\pm 6.3$ mm	$-0.04\pm 0.09$	$-0.001\pm 0.027$
anticipated	$-19.2\pm 11.7$ mm	$+0.21\pm 0.12$	$-0.028\pm 0.024$
maximin	$-4.6\pm 3.9$ mm	$+0.08\pm 0.16$	$-0.054\pm 0.034$

Table VII: Closed-loop comparison against the social concept, paired by layout, direction and repeat (

n=24

pairs of runs each). Positive is worse than social. Uncertainties are the standard deviation of the mean.

Multi-Robot Motion Planning

Multi-Robot Motion Planning coordinates collision-free paths and trajectories for robot teams sharing space, tasks, and changing environments.

See all 145 papers

Mar 5, 2026

Selecting Spots by Explicitly Predicting Intention from Motion History Improves Performance in Autonomous Parking

In many applications of social navigation, existing works have shown that predicting and reasoning about human intentions can help robotic agents make safer and more socially acceptable decisions. In this work, we study this problem for autonomous valet parking (AVP), where an autonomous vehicle ego agent must drop off its passengers, explore the parking lot, find a parking spot, negotiate for the spot with other vehicles, and park in the spot without human supervision. Specifically, we propose an AVP pipeline that selects parking spots by explicitly predicting where other agents are going to park from their motion history using learned models and probabilistic belief maps. To test this pipeline, we build a simulation environment with reactive agents and realistic modeling assumptions on the ego agent, such as occlusion-aware observations, and imperfect trajectory prediction. Simulation experiments show that our proposed method outperforms existing works that infer intentions from future predicted motion or embed them implicitly in end-to-end models, yielding better results in prediction accuracy, social acceptance, and task completion. Our key insight is that, in parking, where driving regulations are more lax, explicit intention prediction is crucial for reasoning about diverse and ambiguous long-term goals, which cannot be reliably inferred from short-term motion prediction alone, but can be effectively learned from motion history.

Long Kiu Chung, David Isele, Faizan M. Tariq +3

Honda Research Institute (HRI), Mountain View, CA. · Georgia Institute of Technology, Atlanta, GA.

Autonomous Parking Autonomous Driving Planning

Ask PDF

Sep 14, 2026

From Learned-Mode AV-Traffic Pairing to Planner Decisions: A Marginal-Preserving Study on Argoverse 2

Joint motion forecasts pair each autonomous-vehicle (AV) future with surrounding traffic, but actor-level metrics do not show whether that structure matters to a planner. We study this question with a marginal-preserving product control that removes AV-traffic pairing among the learned modes while retaining the fixed constant-velocity pair and holding trajectories, actor-level marginals before planner conditioning, candidates, the cost terms and weights, and fallback fixed. The intervention also changes candidate-conditioned concentration. Across twelve runs on 1,400 held-out Argoverse 2 scenarios, the intervention changes 3.0% of route-level offline selections at

τ=4

m. Control-minus-joint recorded-trajectory regret is

-0.026

and

-0.118

at the two training sizes; crossed and seed-

t

intervals span zero. At

τ=1

m, relative costs change in 87.9% of route evaluations and route-level offline selections in 8.1%. Before concentration matching, descriptive outcome estimates favor the control. Most of this gap disappears along an approximate concentration-matching path; the remaining contrasts are

+0.112

and

-0.047

, and both crossed intervals span zero. Actor-level forecast metrics remain identical. Pairing-strength and temperature sweeps show that the decision contrast grows with pairing removal and sharper conditioning. The intervention changes planner decisions even though actor-level metrics remain unchanged. The matching analysis, however, cannot separate any recorded-outcome effect of learned-mode AV-traffic pairing from the accompanying change in conditioned concentration.

Jingyu Wang

Independent Researcher.

Multi-Agent Trajectory Prediction Autonomous Driving Planning

Ask PDF

Jun 14, 2026

JPPD: Joint Prediction_Planning Diffusion with Differentiable Safety Guidance for Dynamic Obstacle Avoidance in Intelligent Transportation Systems

Shared-space transportation operation requires low-speed autonomous platforms to navigate safely and efficiently among pedestrians, service robots, micromobility users, carts, and other road users. Most existing systems decompose this problem into trajectory prediction followed by motion planning, which creates one-way information flow: predicted participant futures influence the robot plan, but the selected robot plan cannot influence the predicted multi-agent evolution. This paper presents a joint prediction-planning diffusion framework that treats participant prediction and robot planning as a single conditional trajectory generation problem, where the model samples the future robot trajectory and all participant trajectories from one coupled distribution using a causal Transformer with cross-trajectory attention. To replace heuristic repulsive post-processing, the framework introduces differentiable safety potential guidance, a time-varying occupancy-probability potential whose gradient directly guides the joint sampler, and conditional flow matching is used to reduce inference steps while preserving multimodal trajectory diversity. The evaluation emphasizes shared-space operational effects, including near misses, blockage time, induced participant deviation, hard-braking events, and embedded latency, rather than treating average displacement error and final displacement error as the main result. Experiments in scenario-grounded simulation, naturalistic pedestrian replay, Isaac Sim validation, and ROS/Orin deployment show that joint sampling improves tail safety and runtime efficiency over a separated prediction-then-planning baseline.

Jiahao Wu, Shengwen Yu

Department of Engineering, Mechanical Engineering, The University of Hong Kong, Hong Kong

Trajectory Generation Diffusion-Based Planning Multi-Agent Trajectory Prediction Dynamic Obstacle Avoidance

Ask PDF

Abstract

Multi-Robot Motion Planning

Explore similar work

Selecting Spots by Explicitly Predicting Intention from Motion History Improves Performance in Autonomous Parking

From Learned-Mode AV-Traffic Pairing to Planner Decisions: A Marginal-Preserving Study on Argoverse 2

JPPD: Joint Prediction_Planning Diffusion with Differentiable Safety Guidance for Dynamic Obstacle Avoidance in Intelligent Transportation Systems