The rollout of robotaxis in London is delayed due to a lack of clear guidance for companies seeking approval. Despite initial plans to launch services later this summer, regulatory and technical hurdles have pushed back the ambitious schedule. Transport for London (TfL) has not yet published the necessary guidance for operators to follow for approval. Both Uber and Wayve have reported that over 100,000 Londoners are on waiting lists for these services. Progress is being slowed while companies await the necessary regulatory framework.
Nav2 has released an update featuring significant new features across several distributions. This includes a new `nav2_ros_common` package for API abstraction and advanced control modes like open-loop control and dynamics support. New components offer capabilities such as Pause and Resume functionality, a Vector Object Server, and an Adaptive Goal Checker. The release also incorporates speedups in the Inflation Layer and improved IPC communication support. Furthermore, a completely redesigned, multi-distribution documentation website has been launched.
The robotics community needs to address the environmental footprint of robots across their lifecycle. The Robotics Eco-Label project aims to quantify robot sustainability by evaluating materials, energy, and design using five metrics to produce an Eco-Score. Other initiatives focus on optimizing purchasing decisions by avoiding overspecification and developing open-source platforms for assessing collaborative robot sustainability. Projects like SolarPeer 360 demonstrate how robotics can help communities achieve energy efficiency and transparent energy sharing. Furthermore, projects like Caretta use robotics for environmental cleanup, highlighting the need for broader education in sustainable practices.
Burns & McDonnell and Gritt partnered to evaluate and deploy AI-powered robotics on utility-scale solar construction sites. Gritt's technology automates labor-intensive tasks by attaching intelligent systems to standard equipment to handle tasks like array assembly and concrete pouring. The collaboration combines Burns & McDonnell's EPC capabilities with Gritt's AI and robotics platform to improve safety, predictability, and efficiency. The technology is specifically designed for outdoor construction environments and continuously learns from real-world field operations. Ultimately, the partnership aims to create tools that empower construction professionals while accelerating the delivery of infrastructure.
The system addresses bottlenecks in Task and Motion Planning (TAMP) by learning symbolic operators directly from demonstration data. It automatically generates macro-operators by discovering causally linked action pairs from the training data. Additionally, the method prunes predicates that are not referenced by any learned operator, reducing the symbolic state. These improvements shorten the effective planning horizon, leading to up to a 4.6x planning speedup across four TAMP domains. This approach enables the system to solve long sequential tasks that baseline methods cannot handle.
A concept-first architecture is proposed for robot spatial understanding, moving beyond metric grid representations. This method uses asynchronous concept agents to build a shared scene graph by instantiating spatial concepts like 'room' and 'door'. Hierarchical constraint propagation guides the detection of features, such as doors, based on room geometry. A complementary prediction-matching loop maintains this structure without needing a pre-existing global metric map. The goal is to achieve a persistent, actionable, and human-interpretable spatial representation for mobile robots.
Large generalist robot policies face issues with reinforcement learning when inference latency causes pauses or jerky movements. The work introduces Asynchronous RL with Intermediate Information (ARLI) to enable RL improvement despite these inference delays. ARLI uses asynchronous inference to hide latency and incorporates state augmentations to restore a near-Markovian structure for the RL policy. Evaluation shows that this framework allows effective finetuning under inference delays, matching or exceeding standard RL performance. This method successfully addresses the incompatibility between RL and latency in deployment settings.
Embodied Agents Systems (EAS) require resilience, which reflects their ability to recover and extend under perturbations, unlike outcome-centric metrics. The authors propose a novel resilience evaluation framework defining metrics for Rebound, Stability, and Graceful Extensibility for EAS execution. This framework systematically evaluates execution processes across 400 household tasks with 10 EAS agents. Metrics-guided optimizations demonstrated diagnostic effects, reducing recovery costs and increasing stability. The research concludes that a trade-off exists among resilience characteristics, necessitating configuration based on deployment requirements.
DreamLedger introduces execution-settled credit files to manage world-model imagination in robot decision loops. It tracks how often predictions are validated against reality, indexed by operating conditions and prediction horizons. Consumed predictions are registered as claims, and outcomes are settled against reality with zero labeling cost, with a credit-gated planning mechanism to reduce wasted imagination. This system demonstrates significant reductions in burned imagination and yields more reliable calibration compared to blind consumption. The framework operates across various model interfaces and remains functional under real-world sensing noise.
Control latents from latent-conditioned adaptive policies are internal representations, limiting closed-loop analysis and model inspection. The work uses conditional flow matching to decode each operational latent into a distribution of quadrotor models for system identification. This decoded distribution enables online predictive tuning of a high-level controller and robustness analysis under disturbances. These methods improve tracking performance significantly under perturbed dynamics and force disturbances. Ultimately, control latents can be converted into physical model ensembles for tuning, diagnosis, and robustness assessment.
The paper proposes safety-aware-stl-mppi, a sampling-based receding-horizon planning framework for robotics. It integrates constraints expressed in Signal Temporal Logic (STL) by encoding them into candidate time-varying control barrier functions (CBFs). These CBFs are then incorporated into a model predictive path integral (MPPI) controller. The method leverages the low computational cost of sampling-based planning while ensuring safety through STL constraints. Experiments on Mars Rover planning and quadcopter control demonstrate the method achieves high safety and efficiency.
Human-robot teaching is increasingly focusing on coupling human instruction with robot learning rather than separate stages. The paper investigates how these coupling dynamics influence teaching effectiveness and human perceptions. To do this, the authors propose a new scale to classify human-robot teaching interactions based on these coupling dynamics. They apply this scale to existing literature to examine the impact of coupling and mental model mismatches on learning outcomes. The research aims to better understand the interplay between human expectations and robot learning through these interaction dynamics.