Robotics & Embodied AI - Weekly Roundup

By modeling policy gradients as Gaussian processes, this work dramatically reduces the sample complexity in reinforcement learning, offering faster convergence and uncertainty estimates at little extra cost.

Forget reinforcement learning; this algorithm learns in real-time without any feedback at all.

LLMs can now intelligently orchestrate multi-agent systems, learning to optimize both individual agent actions and inter-agent cooperation for distributed black-box problems.

Generalist robot policies can achieve 95% success rates on real-world manipulation tasks by continually learning from a fleet of robots, even in the face of distribution shifts and long-tail failures.

Unlock advanced robotic manipulation with FlexiTac, a tactile sensing solution so cheap and easy to integrate, you'll wonder why you were using anything else.

LLMs can now intelligently orchestrate multi-agent systems, learning to optimize both individual agent actions and inter-agent cooperation for distributed black-box problems.

Generalist robot policies can achieve 95% success rates on real-world manipulation tasks by continually learning from a fleet of robots, even in the face of distribution shifts and long-tail failures.

Unlock advanced robotic manipulation with FlexiTac, a tactile sensing solution so cheap and easy to integrate, you'll wonder why you were using anything else.

Physics-informed "grey-box" models aren't just more accurate for structural health monitoring, they can also be greener by reducing computational costs and carbon emissions.

By modeling policy gradients as Gaussian processes, this work dramatically reduces the sample complexity in reinforcement learning, offering faster convergence and uncertainty estimates at little extra cost.

Forget reinforcement learning; this algorithm learns in real-time without any feedback at all.

Simply detecting distribution shifts in visual MBRL is easy; the real challenge is applying the right action-level corrections, which this paper tackles with a novel local expert growth strategy.

AI's non-determinism and data-dependence create critical gaps in the verification, validation, and certification of safety-critical autonomous systems.

Embodied agents can now exhibit coherent, long-horizon, self-directed behavior by reasoning about abstract value trade-offs, a capability previously absent in instruction-following or needs-driven approaches.

The hidden cost of rapidly iterating on AI-enabled perception systems? A growing "Requirements Debt" that threatens auditability, reliability, and certification readiness.

A 48-camera system finally unlocks real-time, room-scale multi-human, multi-robot interaction research in realistic home environments.

Real-time glottis segmentation during Nasotracheal Intubation just got a whole lot faster and more accurate, thanks to a new network that's both lightweight and scale-robust.

Achieve state-of-the-art gait recognition by dynamically fusing body shape and motion features, even when people are wearing coats.

Forget bulky IR cameras: ThermoMesh's tiny thermoelectric mesh can pinpoint sparse heat sources with remarkable sensitivity thanks to clever material science.

Semantic rollouts and town-adversarial regularization can significantly boost zero-shot driving performance in unseen CARLA towns, even without explicit navigation commands or map inputs.

Unlock a baby's-eye view: Reconstruct and replay infant movements on robots to simulate their sensory experiences, offering unprecedented insights into early development.

Unlock the next level of robotic dexterity: this framework lets you co-design robotic hands by optimizing everything from palm structure to fingertip surface curvature.

Adding haptic feedback to robotic surgery training can dramatically improve a surgeon's force regulation accuracy and task efficiency.

Teaching VLMs to "look back" and "look ahead" with lightweight spatial reasoning tasks unlocks surprisingly strong navigation performance.

Achieve clinically relevant accuracy in dynamic bronchoscopy without breath-hold protocols by modeling patient-specific respiratory deformation within a Gaussian splatting framework.

Forget static imitation learning: LaST-R1 unlocks near-perfect robotic manipulation (99.8% success) by adaptively reasoning about physical dynamics *before* acting, then refining with RL.

Forget per-scene optimization: GenWildSplat achieves state-of-the-art 3D reconstruction from sparse, unposed images in real-time using a purely feed-forward approach.

Discovering reusable, semantic "Action Motifs" from human movement data unlocks significant gains in action recognition, motion prediction, and interpolation.

Achieve state-of-the-art open-vocabulary occupancy prediction without any training data, outperforming supervised and self-supervised methods by a large margin.

Control over physical properties like friction and restitution in generated videos is now possible, paving the way for more realistic and controllable video synthesis.

Reconstructing real-world scenes in Minecraft unlocks a customizable embodied AI playground, but only if we can solve the occupancy prediction bottleneck – and this new dataset shows we're not there yet.

Forget painstakingly programming robot interactions – ExoActor uses video generation to hallucinate plausible behaviors, then translates them into robot actions.

Ditch the clunky inverse kinematics: MoCapAnything V2 learns to predict character rotations directly from video, slashing error rates and boosting speed by 20x.

HERMES++ achieves state-of-the-art performance in both future point cloud prediction and 3D scene understanding by unifying these tasks within a single driving world model.

Stop letting SFT ruin your LMMs: PRISM uses on-policy distillation to realign your model *before* RL, boosting performance by up to 6%.

Robots can now better anticipate your actions thanks to a new method that understands the "sub-actions" within your movements.

UAV swarms can now adapt to changing conditions and replan trajectories in real-time by inferring expert-like behaviors from a learned probabilistic world model, avoiding computationally expensive re-optimization.

Hierarchical scene graph matching, learned end-to-end, unlocks fast and accurate robot localization by grounding real-time sensor data against prior architectural maps.

Real-time robot control just got a 50x speed boost thanks to MotuBrain's efficient world action model.

Automated vehicles can achieve fail-operational capabilities by using a hierarchical monitoring framework that combines functional consistency checks with anomaly detection to handle system failures and unfamiliar scenarios.

Marrying short-horizon physics-based control with learned long-horizon intent yields safer and more reliable robot navigation in dense, dynamic environments.

By fusing Bayesian neural networks with Kalman filtering, this work achieves more accurate and robust UAV state estimation than traditional methods in noisy, sparse sensing environments.

Finally, a reinforcement learning algorithm, PGP, can provably find near-optimal policies that respect safety and resource constraints, even when the policy space is non-convex.

Tabular foundation models can dramatically accelerate robot policy learning by enabling efficient global exploration within dynamically constructed policy subspaces.

By pretraining a VLA model with goal-conditioned RL, PRTS learns to reason about goal reachability, leading to substantial gains in long-horizon robotic tasks and zero-shot generalization.

Tackling mean-field control with common noise requires a novel integrated q-function (Iq-function) approach to identify optimal policies as fixed points.

Turns out, language models can reason about mechanical engineering problems, iteratively refining linkage designs by diagnosing failure modes and proposing grounded corrections, all without fine-tuning.

Agentic AI and digital twins can slash traffic light waiting times, outperforming traditional RL methods.

Despite advances in deep learning, manufacturing-focused 3D reconstruction still struggles with reflective surfaces and dynamic environments, highlighting the need for robust hybrid systems.

Existing 3D human mesh recovery systems fall apart for individuals with limb loss, but ResiHMR explicitly reconstructs residual-limb surfaces and performs topology-adaptive optimization, opening the door to more inclusive and accurate human modeling.

Unlock generalist robots by learning manipulation skills directly from the abundance of human activity videos, bypassing the robot data bottleneck.

Nighttime off-road self-driving just got a boost: a new dataset and method robustly handles the dark by fusing infrared and RGB data with a novel memory-attention mechanism.

Self-supervised learning from driving videos can beat fully supervised methods for camera pose estimation, even with orders of magnitude less labeled data.

Cats are helping AI researchers: a Bayesian-inspired model that treats context as a prior significantly improves intent inference for non-speaking agents and avoids shortcut biases.

Forget tedious calibration – DOT-Sim lets you train tactile perception policies in simulation and deploy them directly to real robots with impressive accuracy, thanks to its physically accurate and rapidly calibrated model.

Autonomous driving gets a 30% performance boost in challenging scenarios by having VLAs critique and refine their own driving plans.

Forget sparse, catastrophic rewards – GSDrive uses differentiable 3D Gaussian Splatting to provide dense, physics-based feedback, dramatically improving end-to-end driving policy learning.

Orchestrating autonomous vehicles with dynamic priority scoring in marshaling yards can significantly boost throughput and prevent gridlock compared to static, isolated autonomy.

By representing deformable linear objects as a chain of relative rotations, RopeDreamer achieves state-of-the-art prediction accuracy and topological consistency in long-horizon manipulation tasks.

Quadruped robots can now perform contact-rich manipulation with significantly improved dexterity by learning to "feel" their way through tasks.

Real-time, GPU-accelerated Monte Carlo simulation makes probabilistic safety guarantees for Automatic Emergency Braking systems deployable, not just a validation afterthought.

Kernel Sum-of-Squares optimization can guide sampling-based trajectory optimization out of local minima in high-dimensional contact-rich manipulation tasks.

Finally, a computational model quantifies driving safety by determining if a driver has a collision-free escape route, bridging a 90-year gap between theory and practice.

Autonomous vehicles can now make more judicious lane changes, improving traffic flow and safety, thanks to a federated reinforcement learning system that prioritizes urgency.

Achieve state-of-the-art 3D scene reconstruction from sparse views with 80% less training data by learning to generate, not just match, 3D structures.

Existing methods for controlling actuation-redundant parallel manipulators are flawed, leading to incorrect torque calculations and potentially unstable behavior.

Guaranteeing robot safety in cluttered environments becomes tractable by combining neural radiance fields with reachable set representations for constrained optimal control.

Origami tentacles that deterministically coil and stochastically entangle offer a surprisingly simple and robust solution for universal robotic gripping.

Annotating robot actions just got way faster and more accurate: ATLAS slashes annotation time and error by integrating robot sensor data with video.

Enable robots to adapt to a wide range of disabilities in the workplace with a new persona-based design approach that simplifies the creation of assistive human-robot systems.

Quadruped robots can now nimbly navigate complex 3D terrain using only onboard depth images, thanks to a hierarchical policy that learns strategic navigation and posture adaptation.

Artists can rapidly develop a sense of presence within a robot avatar, opening new creative avenues despite the robot's physical limitations.

A 2D grasp planner can outperform a 3D one in real-world robotic manipulation tasks, while being 25x faster.

Training complex multi-agent RL policies just got 3,500x faster thanks to a new engine that optimizes for memory access and data locality.

LLMs fail over half the time when asked to perform harmful actions in a simulated robotic health attendant setting, even when fine-tuned on medical data.

Achieve real-time robotic action with 79-91% success while generating high-fidelity 4D reconstructions, all within a single unified world model.

VLN agents can navigate more accurately in zero-shot settings by "looking forward, now, and backward," mimicking human navigational strategies.

Rope-assisted climbing robots can now nimbly navigate complex vertical terrains thanks to a new bi-level optimization strategy that coordinates foothold selection and dynamic motion.

Rule-based high-level coaching can drastically improve the safety and sample efficiency of goal-conditioned RL agents in UAV missions, even without pretraining.

Starshaped set filtering slashes computation time and boosts robustness for robot planning in noisy environments, outperforming traditional convex optimization methods.

Robots get a spatial-temporal reasoning boost with STARRY, a world model that aligns future predictions with action generation, leading to a significant jump in manipulation success.

Robots can now navigate complex outdoor environments using only high-level human instructions and readily available GPS/map data, bypassing the need for expensive HD maps or limited short-horizon policies.

Replacing a single skill in a robot's compositional policy can swing task success by up to 50%, and existing behavioral distance metrics can't predict which skill matters most.

Forget photorealism: the next frontier for 3D generation is creating physically plausible, interactive environments that can train robots.

Nighttime UAVs can navigate using only thermal cameras and semantic maps, achieving meter-level accuracy without GPS.

Forget hand-crafted rules and GNN training: LLMs can now autonomously plan robotic tasks, even outperforming human-designed systems.

Achieve robust robot control through intersections by learning phase-conditioned potential functions, sidestepping the instability issues of velocity-dependent methods.

More agents aren't always better: splitting resources too thinly can actually hurt multi-agent system performance, especially when individual agent failure rates increase.

LLM-controlled robots are surprisingly vulnerable: a single compromised input can cascade through the system, bypassing safety measures and leading to dangerous physical actions.

A modified Particle Swarm Optimization algorithm slashes computation offloading latency in vehicular networks, outperforming brute-force methods in dynamic, real-world scenarios.

Semantic SLAM can now understand free-form language queries and ground them in 3D space using only a monocular video feed, opening the door to robots that truly understand and interact with the world around them.

Human motion generation gets a dose of reality: IAM shows that explicitly modeling body morphology and identity leads to more realistic and consistent movements.

Forget replay buffers: TSN-Affinity shows that similarity-guided parameter reuse in TinySubNetworks can achieve strong performance in continual offline RL.

Transformer-based language models don't always win: simpler, TF-IDF-based models surprisingly outperform them in fault localization using industrial bug reports.

RL agents can learn to stay safe in complex environments without excessive conservatism by explicitly modeling and avoiding regions of high uncertainty.

Latent dynamics models like Dreamer can lure you into a false sense of security: their epistemic uncertainty estimates are unreliable because they're biased towards high-reward attractors in the latent space, even when the real world is different.

Guaranteeing zero unsafe state visits during RL training is now possible, opening the door to deploying RL agents in previously inaccessible high-risk environments.

MARL agents can learn to cooperate with unseen partners trained with diverse reward shapings, boosting performance by up to 119% in sparse reward environments.

Achieve a \$1.22M profit in city-scale EV ride-hailing by combining semi-Markov RL with a feasibility-guaranteed MILP projection, outperforming strong baselines and eliminating feeder-limit violations.

Quantum-inspired attention networks can significantly improve task offloading performance in MEC networks, offering a practical path to more energy-efficient and sustainable edge computing.

Neuro-symbolic guidance can dramatically accelerate reinforcement learning in sparse-reward environments, even when the symbolic knowledge is imperfect.

Randomly sampling tasks in offline RL hurts zero-shot generalization, but extracting task vectors directly from the dataset boosts performance by 20%.

Evolving interpretable control policies for multi-task robots is now possible: MATPG leverages genetic programming to create a single agent that masters multiple continuous control tasks.

LLMs fail to grasp basic spatial concepts and cultural nuances encoded in demonstratives like "this" and "that," revealing a surprising lack of embodied cognition despite their vast training data.

Automating system-level testing for distributed robotics is now more practical with a new language that handles complexity, non-determinism, and dynamic reconfiguration.