Social Mini-Games, or SMGs, are the deceptively simple moments of friction that plague robot deployments everywhere. Two delivery robots converging on the same doorway, a fleet of mobile robots meeting head-on in a narrow hallway, several autonomous vehicles arriving simultaneously at an unregulated intersection: these are the scenarios the survey formalizes. The authors define an SMG precisely as a situation in which the preferred or optimal trajectories of two or more agents collide within a common time interval, typically lasting long enough that someone must yield. What distinguishes these encounters from ordinary navigation is what the researchers term “agency.” In SMGs, small actions carry outsized consequences: unless at least one agent alters its course, a deadlock or collision becomes inevitable. The paper contrasts this with general multi-robot navigation, where agents operate with low to medium agency and their trajectories remain largely decoupled. Crucially, the authors note that humans resolve these conflicts gracefully, modulating their velocity by just enough to let another person pass through a doorway first while barely deviating from their preferred speed, a behavior that remains remarkably difficult for robots to replicate.

The mathematical machinery underpinning the survey is substantial. The authors model multi-robot navigation as a partially observable stochastic game, in which each robot maintains a state combining visible parameters, such as position and velocity, with hidden internal parameters like preferred speed and heading. Robots generate local observations through their sensors, follow control-affine dynamics, and incur running costs based on goal distance, control effort, and deviation from preferred paths. Collisions are formalized through the overlap of convex hulls inflated by safety margins. To quantify when a general navigation scenario tips into an SMG, the researchers introduce a geometric “liveness function” that measures the angle between relative displacement and relative velocity vectors. When two agents approach a capacity-one resource like a doorway with nearly parallel position and velocity vectors, the liveness value approaches zero and perfect symmetry, and thus contention, ensues. Without an explicit symmetry-breaking mechanism, such interactions converge to undesirable equilibria: deadlocks or oscillatory behavior in which neither robot makes progress.

Beyond formal definitions, the survey catalogs a canonical set of SMG scenarios drawn from real-world topologies: doorways, intersections, two-way hallways, L-corners, blind corners, and crowded, parallel, perpendicular, and circular traffic patterns. Each scenario activates a distinct subset of SMG characteristics, exposing different deadlock mechanisms, visibility challenges, and coordination demands. Alongside these scenarios, the paper proposes a battery of evaluation metrics that go well beyond the standard measures of velocity smoothness, delay, and path deviation. Because SMG agents possess high agency, the authors argue, evaluation must capture game-theoretic outcomes. They introduce flow rate, which quantifies density-normalized throughput through a spatial bottleneck; a fairness metric rooted in game theory that measures the gap between each agent’s realized reward and what it would earn under a fully cooperative optimum, with weighted priorities allowing, for example, an emergency robot to justifiably claim the right of way; and an influence score computed through counterfactual analysis, measuring how much one agent’s presence perturbs the control behavior of all others by re-solving the game with that agent removed.

The centerpiece of the work is a taxonomy that organizes existing SMG solvers across multiple design dimensions. At the paradigm level, the authors sort methods into four broad families. Multi-agent reinforcement learning approaches, including value-based methods like Q-learning, multi-agent deep Q-networks, and policy-based frameworks such as MADDPG, treat an SMG as a localized stochastic subgame that emerges wherever non-stationarity plagues an agent’s learning process. Multi-agent path finding methods, including Conflict-Based Search, Increasing Cost Tree Search, and Prioritized Planning, resolve conflicts on discretized graphs but classically assume centralized, fully observable, offline computation. Optimization-based approaches, spanning reciprocal velocity obstacles, model predictive control, and control barrier functions, cast navigation as constrained optimal control in which shared inequality constraints activate among coupled agents. A fourth category collects heuristics, evolutionary methods, and hybrid systems such as genetic-algorithm-based topological optimization for warehouse logistics.

The taxonomy’s finer axes reveal the design trade-offs that define the field. Coordination can be centralized, with a single authority issuing commands; decentralized, with each agent deciding independently from local observations; or distributed, combining centralized training with communication-aware decentralized execution. Communication may be present, allowing agents to share plans and intentions, or absent, forcing reliance on shared norms. Deadlock handling splits into proactive prevention, which plans ahead like a chess player thinking several moves deep, and reactive resolution, which waits for a stall to emerge and then breaks it using rules like the right-hand rule, auctions, or adaptive rotational strategies. The authors also formalize “invasiveness,” a property describing how dramatically a solver perturbs an agent’s preferred trajectory: minimally invasive solvers merely tweak speed without altering course, while invasive ones may force an agent to stop entirely and wait. Cooperation ranges from fully cooperative settings with shared cost functions, exemplified by methods like PRIMAL, to non-cooperative settings where agents pursue private objectives, and observability spans full knowledge of all agents’ states down to minimal awareness in which other robots are treated as static obstacles.

To make these abstractions concrete, the survey profiles representative solvers that have become de facto baselines. CADRL uses deep reinforcement learning with LSTM networks to compress variable numbers of nearby agents into fixed-size observations, producing smooth learned collision-avoidance policies, though without formal safety guarantees. The Right-Hand-Rule method enforces clockwise yielding priorities, and can be implemented through safety barrier certificates computed as quadratic programs, but its rigid ordering can violate fairness constraints. Auction-based approaches such as SocialMAPF let agents bid on priority using control barrier functions, balancing efficiency against occasionally longer travel times. IMPC-DR employs infinite-horizon model predictive control over modified buffered Voronoi cells with warning bands, solving a convex optimization at every replanning step and proving that no stable deadlocks can occur under specified conditions, all while relying only on local communication. Perhaps most striking is ORCA-MAPF, a hybrid in which agents normally navigate with the linear-programming-based ORCA velocity scheme but switch to a locally confined, grid-based path-finding solver, agreed upon deterministically without a central controller, whenever a deadlock is detected; in some tests this lifted navigation safety rates from 15 percent to 99 percent.

Underlying all of these approaches are five properties the authors identify as desirable: provable safety, achievable through control barrier functions that guarantee forward invariance of safe sets or through velocity-obstacle frameworks solvable as linear programs; liveness, the guarantee that robots keep making progress and never freeze in place; welfare maximization, which balances individual priorities against collective goals, such as granting an ambulance robot priority at an intersection; social compliance, the ability to respect human norms and expectations; and scalability, which the paper argues should be measured against the size of the active coupling set of interacting agents rather than the total robot population, since many solvers handle large fleets comfortably until a bottleneck triggers a localized, tightly coupled game.

The survey does not shy away from open problems. Visual-input navigation, in which robots must guarantee safety and liveness directly from raw lidar point clouds and camera feeds without expensive state estimation, remains computationally punishing, as evaluating barrier functions and their derivatives in real time strains onboard hardware. Navigation alongside humans raises questions about whether trajectory-forecasting models that excel on sparse crowd datasets transfer to tightly coupled SMG interactions involving passing, weaving, and yielding. And the authors find current simulation environments wanting: most model only single agents in open spaces, and few, among them SocialGym and SEAN 2.0, approach the configurability needed to train socially compliant robots in geometrically constrained settings. To accelerate the field, the team has open-sourced SMGLib, an extensible simulation environment that lets researchers run planners through canonical SMG scenarios, log metrics, and visualize trajectories in real time.

The authors’ recommendations are pointed: the community should prioritize non-invasive techniques that let agents navigate without dramatic deviations, and should favor deadlock prevention over after-the-fact resolution, since avoiding stalls from the start yields smoother and more reliable motion. Drawing an explicit parallel to the influential 2019 survey that unified and streamlined multi-agent path-finding research, Chandra and colleagues express the ambition that their taxonomy will do the same for social mini-games, giving newcomers a map of a once-fractured territory and giving practitioners a common language for matching algorithms to the doorways, hallways, and intersections where robots must now earn the right of way.

Trajectory prediction is critical for autonomous vehicles, allowing them to anticipate the future movements of other agents on the road to ensure safety and smooth navigation. Traditional methods operate under the assumption of zero latency—an idealized scenario where predictions are made instantaneously without delay. However, in practical autonomous driving systems, latency is unavoidable due to the time required for data sensing, processing, and prediction calculations. The conventional oversight of this latency leads to predictions that are already outdated the moment they are generated, resulting in reduced accuracy and potential safety risks.

What sets the LatenAux framework apart is its strategic acknowledgment and incorporation of latency within the prediction process. Rather than viewing latency as a mere hindrance, the researchers reconceptualize it as auxiliary contextual information that can be leveraged to improve forecasts. This reconceptualization manifests through a dual-task learning structure that separates prediction into two interconnected branches: a primary task tasked with forecasting trajectories within a valid future horizon, and an auxiliary task dedicated to interpreting latency-inclusive observational data.

This auxiliary branch stands as a novel innovation, ingesting inputs reflective of the latency period—data that previous models discarded or ignored. By doing so, LatenAux embraces the typically “stale” latency data and employs it as valuable auxiliary knowledge. A core feature of this architecture is a progressive feature alignment strategy, which facilitates the transfer of latency-aware insights from the auxiliary branch to the primary prediction branch. This approach ensures that the primary model internalizes nuanced latency cues, enabling it to produce more accurate trajectories without needing explicit latency information at inference time.

Furthermore, LatenAux distinguishes itself through the introduction of a soft feature-consistency mechanism that governs how auxiliary information from the latency-inclusive branch influences the primary branch. Unlike harsh constraints that may overfit or restrict learning, this mechanism gently aligns feature representations across both scene context and state query levels. This balance enriches the internal feature space, promoting robustness while mitigating risks of learning degradation often associated with direct feature constraints.

Complementing this, auxiliary queries generated from latency-affected observations serve as informative priors for the primary prediction branch. These priors supply contextual guidance, helping to refine and calibrate the trajectory predictions in a dynamic and adaptable manner. This synergy between primary and auxiliary components forms the backbone of LatenAux’s superiority over existing state-of-the-art models.

The efficacy of LatenAux has been demonstrated through exhaustive experiments on two extensive, real-world autonomous driving datasets. These datasets, featuring complex urban driving scenarios with diverse agent behaviors, provided an ideal proving ground for the model’s performance. The results consistently showed that LatenAux not only enhances latency-aware modeling capabilities but also delivers trajectory predictions that are significantly more precise and dependable compared to traditional latency-agnostic approaches.

Importantly, the adaptability of LatenAux across a range of latency durations marks a major advance in practical applicability. Autonomous driving systems vary widely in hardware capabilities and system configurations, resulting in differential latency profiles. LatenAux’s inherent flexibility ensures that autonomous systems equipped with varied specifications can uniformly benefit from latency-aware forecasting, turning a fundamental limitation into a valuable feature.

Professor Haiyang Yu, leading the research team, emphasizes the revolutionary shift this framework brings: “Our latency-aware trajectory prediction framework opens a fundamentally different pathway toward practical trajectory forecasting. By explicitly addressing latency, we provide a new direction that can bridge the gap between theoretical models and real-world deployments.” This vision could catalyze the development of safer, more intelligent autonomous vehicles equipped to handle the inherent delays in their sensing and computational subsystems.

Ph.D. candidate Zhengxing Lan, who played a significant role in validating the approach, notes, “Through extensive experimental validation, LatenAux has demonstrated its clear advantage. Its ability to incorporate latency as auxiliary knowledge not only boosts prediction accuracy but also underpins the reliability of trajectory forecasts essential for downstream planning modules.”

Lingshan Liu, another key contributor, reflects on the broader implications: “The demonstrated adaptability of our model effectively converts a core technological constraint into an operational strength. By enhancing robustness across system variances, LatenAux ensures that autonomous driving platforms remain effective under diverse and realistic scenarios, accelerating the pathway toward widespread adoption.”

This work also signals a broader shift in autonomous systems design philosophy—embracing real-world imperfections such as processing latency and utilizing them proactively, rather than marginalizing or ignoring them. The LatenAux framework may pave the way for other domains to incorporate auxiliary learning paradigms that exploit system limitations to improve overall performance and reliability.

Published in the reputable Communications in Transportation Research, the study benefits from the journal’s rigorous peer review and position as a leading venue for cutting-edge transportation research. With an impact factor rising to 14.5 in 2024 and a top ranking in the transportation category globally, publication in this journal underscores the significance and quality of this contribution.

This breakthrough has the potential to transform how autonomous driving systems manage temporal delays, providing a robust foundation for next-generation trajectory forecasting. As autonomous vehicles inch closer to full deployment, innovations like LatenAux will be critical in ensuring their safe and reliable operation amidst practical constraints.

For researchers, engineers, and policymakers aiming to integrate autonomous systems into everyday transportation, the findings of this study highlight the importance of reconsidering latency not merely as an obstacle but as an opportunity to enhance predictive intelligence and safety.

Subject of Research: Latency-Aware Trajectory Prediction for Autonomous Driving

Article Title: LatenAux: Towards Latency-Aware Trajectory Prediction for Autonomous Driving via Consolidated Auxiliary Learning

News Publication Date: 31-Mar-2026

Web References:

• https://doi.org/10.26599/COMMTR.2026.9640010
• https://www.sciopen.com/journal/2097-5023

References:
Yu, H., Lan, Z., Liu, L., et al. (2026). LatenAux: Towards Latency-Aware Trajectory Prediction for Autonomous Driving via Consolidated Auxiliary Learning. Communications in Transportation Research. DOI:10.26599/COMMTR.2026.9640010

Image Credits: Communications in Transportation Research

Keywords

Latency-aware prediction, autonomous driving, trajectory forecasting, auxiliary learning, feature alignment, latency-inclusive observations, progressive feature transfer, trajectory accuracy, deep learning, autonomous vehicle safety, robust forecasting, intelligent transportation systems

The ability to perceive hidden areas has long been a challenge for robotic systems, especially in complex environments. HoloRadar employs radio waves to create a multifaceted view of a scene that exists beyond direct visual contact. Unlike traditional imaging methods that depend on visible light, which can be unreliable in low-light conditions or obstructed views, HoloRadar operates effectively in a diverse array of lighting conditions. This unique capability offers significant potential to enhance safety measures for robots and driverless cars, allowing them to make more informed decisions in real time.

What sets HoloRadar apart is its novel approach to interpreting radio signals, characterized by an astounding observation: the lengthy wavelengths of radio waves, which are often deemed a disadvantage in imaging, can be advantageous in discerning hidden objects. This phenomenon transforms ordinary surfaces, such as walls and ceilings, into reflective mediums. Instead of merely passing through barriers, radio waves can bounce off these surfaces, gathering crucial data about the hidden locations and providing robots with a clearer understanding of their surroundings.

At the heart of HoloRadar’s functioning lies a sophisticated AI system that processes the radio signals reflected back to the sensor. When a single radio pulse is transmitted, it ricochets multiple times off various surfaces before returning, creating a complex pattern of reflections that would normally confound traditional analysis methods. To handle this complexity, the research team devised a two-phase processing approach that combines machine learning with physics-based modeling to unravel these intricate reflections.

During the initial phase, HoloRadar conducts a resolution enhancement of the raw radio signals and discerns multiple returns, enabling the system to identify various paths the signals have taken. Once this information is gathered, the second phase utilizes a physics-guided model to trace these signals back to their origins, effectively reversing the mirror-like effects of the environment. This step is critical as it allows the system to reconstruct the actual three-dimensional layout of the scene and to pinpoint the locations of hidden objects and obstacles.

The researchers have conducted experiments that demonstrate the capabilities of HoloRadar in real-world settings. Tests performed on mobile robots navigating indoor spaces, such as hallways and building corners, have shown promising results. The system successfully achieved the reconstruction of environments, accurately identifying walls, corridors, and even human subjects located outside the robots’ direct lines of sight. This remarkable functionality signifies a substantial advancement in how robots perceive and interact with their environments.

HoloRadar boasts a pioneering design tailored for operational versatility in a wide range of scenarios. This technology is unlike previous attempts to achieve NLOS vision, often constrained by heavy and cumbersome scanning equipment, which limits practical applications. Conversely, HoloRadar’s mobility and real-time processing capabilities make it adaptable to the diverse environments robots encounter daily, including both indoor and outdoor settings.

Moreover, HoloRadar doesn’t aim to replace existing sensor systems but rather to enhance them. Many autonomous vehicles already utilize LiDAR technology for direct object detection. By incorporating HoloRadar into their toolkit, robots can extend their perceptual reach, gaining essential insights about potential hazards that may not be visible through conventional sensors. This multifaceted approach not only augments existing safety measures but also offers robots increased response time against dynamic threats.

The implications of HoloRadar stretch beyond just enhanced safety for autonomous machines. As researchers aim to deploy this technology outdoors in the future, they envision a myriad of use cases, from emergency response robots navigating disaster-stricken areas to delivery drones efficiently avoiding obstacles. Realizing the long-term potential of HoloRadar could revolutionize urban planning, transportation, logistics, and many other industries reliant on robotics.

As the development of HoloRadar progresses, ongoing research will focus on accommodating diverse environments, tackling challenges such as increased distances and variable outdoor conditions. The long-term goal is to facilitate an advanced understanding of surroundings for robots, allowing them to maneuver intelligently through complex environments as humans do. By integrating this technology into robotics, the team aims to contribute significantly to advancements in autonomous navigation and situational awareness.

This constellation of innovations arises from extensive experimentation at the Wireless, Audio, Vision, and Electronics for Sensing (WAVES) Lab at the University of Pennsylvania’s School of Engineering and Applied Science. Funded by the university itself, this research signifies a significant investment in the future of robotics and artificial intelligence, emphasizing the institution’s commitment to pioneering exploration in this rapidly evolving field.

The release of HoloRadar not only signifies an important step in robotics but also raises intriguing questions about the future of machine perception. As researchers continue to push the limits of what is possible, the world is undoubtedly on the cusp of a technological evolution that will change how we interact with machines and, ultimately, each other.

With the introduction of HoloRadar, the vision of robots that perceive their environments with the same acumen as human beings draws closer. By tapping into the untapped potential of radio waves and applying cutting-edge AI technologies, the engineers at Penn are not merely developing tools; they are laying the groundwork for a safer, smarter world powered by interconnected autonomous systems.

Subject of Research:
Article Title: Non-Line-of-Sight 3D Reconstruction with Radar
News Publication Date: 4-Dec-2025
Web References: HoloRadar Project
References: N/A
Image Credits: Credit: Sylvia Zhang, Penn Engineering

Keywords

HoloRadar, Non-Line-of-Sight Vision, Radio Waves, Robotics, AI, Autonomous Vehicles, 3D Reconstruction, Safety Technology, Machine Learning

At the heart of GNSS positioning lies the integration of satellite signals with correction data to mitigate atmospheric errors. Precise Point Positioning (PPP) has long been the gold standard, delivering centimeter-level accuracy by leveraging correction data for satellite orbit, clock, atmospheric delays, and more. However, traditional PPP’s relatively slow convergence time limits its applicability in time-sensitive scenarios. To address this hurdle, Precise Point Positioning–Real-Time Kinematic (PPP-RTK) techniques have emerged, complementing PPP by incorporating real-time atmospheric corrections that accelerate ambiguity resolution and shorten convergence periods. Despite these advances, PPP-RTK’s efficacy is critically dependent on the quality of atmospheric correction data, which is notoriously sensitive to fluctuations in satellite elevation angles, geomagnetic activity, solar influences, and the spatial configuration of ground stations.

Such susceptibility has been a persistent thorn, as atmospheric disturbances can introduce errors on the order of centimeters to decimeters, compromising real-time positioning accuracy. Traditional approaches to assessing atmospheric correction quality have relied heavily on empirical models derived from extensive historical datasets or on dense networks of dedicated monitoring stations. These frameworks, while useful, often hinder adaptability to dynamic conditions and limit the scalability of GNSS augmentation services. Recognizing this gap, the research team embarked on developing a self-reliant, scalable method capable of delivering real-time quality assessments without dependency on external validation points or legacy data.

Their solution harnesses the statistical robustness of leave-one-out cross-validation (LOOCV), a technique conventionally rooted in machine learning and statistical inference, now repurposed for atmospheric correction validation within GNSS networks. This approach cyclically designates each individual reference station within a network as a “validation point,” while leveraging the remaining stations to generate the correction dataset. By systematically rotating through all stations as test cases, the method internally evaluates the fidelity of atmospheric corrections in a fully dynamic, data-driven manner. This innovative internal validation framework yields real-time quality metrics that convey the reliability of corrections across the spatial grid of monitoring stations.

Crucially, the study incorporates these quality metrics directly into the PPP-RTK service output, broadcasting alongside traditional correction data. This paradigm shift empowers end-users to not only receive atmospheric corrections but also instantly gauge their accuracy and stability. Such transparency is a game-changer, especially in safety-critical and scientifically demanding contexts where confidence in the navigation solution’s integrity is indispensable. The capability to access correction quality in real time equips users to dynamically adapt to uncertain atmospheric conditions, ensuring operational continuity and precision.

Experimental validation spanned two distinct atmospheric environments: a stable, mid-latitude European network comprising 21 stations and a low-latitude, ionosphere-affected Hong Kong network with 19 stations. Results revealed remarkable stability in tropospheric corrections, maintaining accuracy within approximately 2 centimeters, while ionospheric corrections exhibited variability ranging from 2 to 15 centimeters contingent on solar activity levels. Impressively, over 90% of the quality estimates corresponded closely with observed error deviations, affirming the method’s reliability in diverse geophysical and geomagnetic contexts.

In the applied realm, the implementation of LOOCV-driven quality monitoring translated into tangible improvements in PPP-RTK positioning performance. The method facilitated enhancements in positioning accuracy by a notable margin—ranging from 6 to 29 percent in Europe and 9 to 20 percent in Hong Kong networks. Beyond accuracy, convergence times saw accelerated reductions, an outcome with profound implications for real-time navigation systems that must rapidly establish precise locations. Perhaps most strikingly, even amid intense geomagnetic storms—when ionospheric disturbances are at their peak—the method sustained positioning improvements up to 40%, underscoring its robustness under challenging space weather conditions.

Professor Xingxing Li, the study’s corresponding author, emphasized the transformative potential of this approach, noting that embedding self-monitoring within GNSS correction services liberates the system from reliance on supplementary ground infrastructure or rigid empirical models. The intrinsic adaptability embedded in the leave-one-out cross-validation framework enables seamless operation across various network densities and environmental conditions, marking a new era of self-sustaining GNSS augmentation. Prof. Li further highlighted the critical safety dimension, pointing out that autonomous systems and disaster response mechanisms stand to benefit immensely from a navigation solution that transparently communicates correction reliability amidst ever-changing atmospheric phenomena.

From a broader perspective, this advancement addresses a long-standing bottleneck in satellite navigation—namely, the real-time appraisal of correction integrity. The integration of dynamic quality information into PPP-RTK services sets a precedent for future augmentation architectures, fostering increased trustworthiness required for next-generation applications including intelligent transportation, precision agriculture, infrastructure surveying, and even spaceborne platforms. The research anticipates seamless assimilation of this methodology into global satellite-based augmentation systems, accelerating widespread adoption and enhancing navigational reliability worldwide.

Furthermore, this novel approach holds profound implications during periods of heightened solar and geomagnetic activity, which historically have introduced severe positioning errors due to intensified ionospheric disturbances. By maintaining centimeter-level accuracy and delivering reliable uncertainty bounds in such volatile conditions, the method safeguards critical infrastructure and operational processes that depend on uninterrupted, accurate location data. This capacity could redefine operational protocols across sectors vulnerable to space weather impacts, enhancing resilience and ensuring continuity.

The study’s findings validate an emergent paradigm—moving from correction delivery as a black-box service to a transparent, self-evaluating system that serves users with rich, actionable information about correction quality. By empowering navigation solutions with self-assessment capabilities, GNSS technologies can transcend existing limitations, embracing complexities intrinsic to earth-space signal propagation. This transition not only paves the way for improved positional accuracy but also fosters a culture of awareness and trust within the satellite navigation ecosystem, which is essential as autonomous systems become increasingly interwoven with daily life.

In summary, the introduction of leave-one-out cross-validation as a real-time quality monitoring tool for grid-based atmospheric corrections in GNSS PPP-RTK services represents a watershed moment in satellite navigation. It circumvents previous dependency on external validation infrastructure, offers robust performance across atmospheric variabilities, and dramatically uplifts positioning accuracy and convergence speeds. As GNSS applications permeate ever more critical domains, the ability to trust and verify atmospheric corrections on the fly will be indispensable to the next generation of navigation technologies.

Subject of Research: Navigation

Article Title: Quality monitoring of grid-based atmospheric corrections in GNSS PPP-RTK service using leave-one-out cross-validation

News Publication Date: 22-Sep-2025

References:
DOI: 10.1186/s43020-025-00178-5

Image Credits: The authors

Keywords: GNSS, Precise Point Positioning, Real-Time Kinematic, atmospheric corrections, leave-one-out cross-validation, positioning accuracy, ionospheric disturbances, tropospheric corrections, satellite navigation, solar activity, geomagnetic storms, navigation reliability