Sign in Subscribe
Robotics

🤖 Robotics Pulse · 2026-06-19 00:02 UTC

ROBOTICS PULSE

Friday, June 19, 2026

Your daily briefing on robotics and AI from official and peer-reviewed sources.

⚡ TL;DR

Qwen-RobotManip and a wave of companion VLA papers demonstrate that alignment-driven scaling is rapidly closing the gap between language-model generalization and real-world robot dexterity. Today's feed is exceptionally dense: 40-plus cs.RO papers drop alongside DARPA physical-intelligence moves and fresh NIST AI governance proofs, signaling a field accelerating on all fronts simultaneously.

🤖 ROBOTICS

HUMANOID LOCO-MANIPULATION

  • HALOMI learns humanoid loco-manipulation with active hand-eye coordination directly from human demonstrations, bypassing the need for a precise world-frame tracking controller. [1]
  • HumanoidArena introduces an egocentric hierarchical benchmark that couples high-level task decisions with whole-body dynamic execution for humanoid policy evaluation. [2]
  • Mobile Pedipulation uses a nonlinear model predictive controller built on a three-rigid-body model to let a wheeled bipedal robot slide objects across flat surfaces using its wheeled legs. [3]

VLA MODEL PROLIFERATION

  • Qwen-RobotManip applies the language-model alignment-then-scale recipe to robotic manipulation, reporting genuine cross-task generalization in a detailed technical report released June 16. [4]
  • ThinkingVLA interleaves vision and language reasoning chains inside a VLA backbone, improving performance on long-horizon tasks that demand spatial anticipation. [5]
  • PearlVLA refines action plans progressively in latent space, targeting the speed-versus-deliberation tradeoff common to current VLA architectures. [6]
  • MuseVLA augments a standard RGB-only VLA with adaptive multimodal sensing, addressing physical properties that cameras alone cannot resolve. [7]
  • Motion-Focused Latent Action extracts latent action labels from egocentric human videos, enabling VLA pre-training on abundant human footage that lacks robot action annotations. [8]
  • Object-Centric Residual RL adds a zero-shot sim-to-real residual policy on top of a frozen VLA, reducing execution errors in precise physical interactions without re-collecting real data. [9]
  • WeaveLA uses event-driven cross-subtask latent memory to pass state information across sub-task boundaries in repetitive manipulation sequences. [10]
  • GeneralVLA-2 adds geometry-aware 3D reconstruction and governed episodic memory to an existing hierarchical VLA planner. [11]
  • WAM-RL extends World-Action models beyond expert demonstrations using reconstruction rewards and online video supervised fine-tuning. [12]
  • LAGO Policy addresses jerk and collision in asynchronous diffusion-policy deployment through latency-aware chunking and goal-directed collision-free planning. [13]
  • VERITAS proposes a generator-verifier framework enabling a generalist robot policy to steer itself at inference time and autonomously improve from its own experience. [14]

DEXTEROUS MANIPULATION AND GRASPING

  • DexLink Hand achieves 16 degrees of freedom in a human-scale form factor using linkage-driven transmission, targeting the dexterity-compactness-affordability tradeoff. [15]
  • Do as I Do extracts dexterous manipulation data from everyday human videos by estimating hand-object interaction, providing scalable training signal for multi-fingered hands. [16]
  • Zero-Shot Long-Horizon Dexterous Manipulation grounds language instructions into executable 3D task plans from calibrated multi-view RGB images without training an end-to-end policy. [17]
  • GCNGrasp-VP plans views to reveal occluded task-relevant regions before attempting task-oriented grasps, integrating affordance reasoning with active perception. [18]
  • EAGG conditions a grasp generator on geometry-aware graph embeddings, enabling generalization across parallel grippers and dexterous end effectors from a single model. [19]
  • TactSpace learns a physics-enriched shared latent space to transfer tactile sensing policies from simulation to real sensors, attacking the sim-to-real gap in contact-rich manipulation. [20]
  • HT-Bench introduces a benchmark for full-hand tactile representation learning anchored to egocentric vision. [21]
  • WireCraft releases a simulation benchmark for industrial deformable linear object manipulation, covering wire-routing and cable-threading tasks. [22]

PERCEPTION, SLAM, AND NAVIGATION

  • FAST-LIVGO fuses LiDAR, inertial, visual, and GNSS data with degeneracy-robust switching to maintain centimeter-level accuracy across long-range, highly dynamic environments. [23]
  • RICH-SLAM builds a continuous Hilbert-space occupancy map incrementally from sparse radar returns, improving radar SLAM fidelity in adverse weather. [24]
  • Sensor Configuration Matters provides the first systematic comparison of visual-inertial SLAM configurations on quadruped robots across diverse real-world terrains. [25]
  • C-ARC delivers real-time adaptive range clustering for non-repetitive LiDAR sensors, whose irregular scan patterns break methods designed for spinning sensors. [26]
  • High-Accuracy Event-Based Underwater SLAM replaces degraded time-surface methods with a stereo event-camera pipeline robust to fluctuating velocity and repetitive underwater texture. [27]
  • Spatially Stratified Distillation bridges heterogeneous radar hardware by distilling dense spinning-radar maps into a form queryable by low-cost 4D automotive radar. [28]
  • Viking Hill Dataset releases co-registered LiDAR, 4D imaging radar, and camera data with per-tree annotations across seasonal conditions for forest robot perception. [29]
  • OneCanvas projects multi-view patch features onto a single equirectangular panoramic canvas for efficient 3D scene understanding in VLMs without task-specific geometry encoders. [30]
  • FLAP plans UAV trajectories under strict FOV and sensing-range constraints in unknown cluttered 3D environments without a prior map. [31]
  • Monocular 3D Occupancy Perception uses hybrid 2D-3D learning to predict sidewalk occupancy for delivery robots and wheelchairs, where road-centric pipelines fail. [32]
  • ERQA-Plus releases a diagnostic benchmark requiring embodied agents to reason about spatial relations, procedures, human intentions, and commonsense consequences. [33]

QUADRUPED AND LEGGED ROBOTS

  • ZiMPedance introduces impedance-aware zero-moment-point modeling to handle oscillatory forces introduced by passive spring-based payload arms on quadruped carriers. [34]
  • Damage Adaptation in Seconds demonstrates proprioceptive recovery from catastrophic damage in soft-actuated architected-material systems in under one minute. [35]

SOFT ROBOTS AND NOVEL HARDWARE

  • Selective Unit-Cell Actuation moves actuator-lattice co-design to the unit-cell level in soft lattice structures, enabling distributed morphology control across the robot body. [36]
  • Shape Sensing of Continuum Robots uses Direct Laser Writing to fabricate integrated strain sensors directly onto continuum surgical robot bodies for real-time shape reconstruction. [37]

AUTONOMOUS VEHICLES AND DRONES

  • TerraTransfer learns end-to-end driving policies without any expert demonstrations by combining terrain-aware transfer and self-supervised objectives. [38]
  • Autonomous Emergency Braking annotation system uses machine learning to flag rare delayed and false AEB triggers that comprise less than 5 percent of thousands of daily events. [39]
  • A Mixed-Reality Hardware-in-the-Loop testbed integrates physical mobile robots with a high-fidelity simulator for validating AV stacks in safety-critical edge cases. [40]
  • CABLE cuts V2X communication overhead by compressing video tokens cloud-side before prefill, enabling open-vocabulary LMM perception at the network edge. [41]
  • ED3R coordinates energy-aware robotic agents for cooperative wildfire and disaster detection under strict resource and uncertainty constraints. [42]
  • Accountability in Autonomous Drone Firefighting examines how integrating autonomous drones into emergency teams shifts established accountability relationships based on a live field trial. [43]
  • DroneLets formalizes repeatable drone-emergency-service collaboration patterns to replace ad hoc coordination in first-responder workflows. [44]

HUMAN-ROBOT INTERACTION AND TELEOPERATION

  • Generating Natural Robot Gestures applies iterative reinforcement learning with human feedback via LLMs to produce expressive pointing and communicative motions for the Pepper humanoid. [45]
  • Arm Kinematic Correction uses deterministic geometric inference to recover accurate limb poses from a single RGB-D camera even under self-occlusion during teleoperation. [46]
  • Engagement-Aware Human-in-the-Loop proposes keeping users actively involved throughout robot operation rather than calling for intervention only at failure. [47]
  • Continual Online Personalization of Exoskeleton Control uses manifold-aware experience replay to adapt gait assistance in real time without catastrophic forgetting across locomotor tasks. [48]

SPACE ROBOTICS

  • DARPA confirmed the Robotic Servicing of Geosynchronous Satellites technology is on track to launch in 2026, representing the most significant milestone yet for the RSGS program. [49]

SIMULATION AND DATA INFRASTRUCTURE

  • MagicSim provides a unified execution substrate linking robot control, skills, and planning within a single simulation layer, eliminating disconnected "magic" action shortcuts. [50]
  • GASE uses Gaussian Splatting to automatically reconstruct high-fidelity simulation environments from real-scene captures, minimizing the sim-to-real appearance gap. [51]
  • AnnotateAnything automatically enriches raw 3D simulation assets with semantic, interactive, and physical annotations needed for robot manipulation policy training. [52]
  • EBench diagnoses generalist mobile manipulation policies across 26 tasks annotated along 5 capability and 4 generalization dimensions, going beyond single success-rate scores. [53]
  • Qwen-RobotNav releases a scalable navigation base model whose observation strategy is reconfigurable at inference time to serve instruction-following, object search, and autonomous driving from one backbone. [54]

🧠 AI & MODELS

REASONING AND PLANNING

  • MIT researchers showed that a small AI model trained to ask targeted questions in a Battleship-style game outperforms large frontier models at the same task for roughly 1 percent of the compute cost. [55]
  • Diffusion-Proof applies masked diffusion as an alternative to autoregressive generation for formal theorem proving in Lean, claiming improved coverage of proof search space. [56]
  • Fixed-Point Reasoners introduce convergence regularization for looped Transformers so iterative deep computation stabilizes rather than diverges on compositional reasoning tasks. [57]
  • Looped World Models apply looped architectures to world-model simulation, resolving the tension between depth-of-computation and deployment cost for long-horizon rollouts. [58]

REINFORCEMENT LEARNING AND ALIGNMENT

  • RODS synthesizes new multi-turn tool-use training data online, targeting tasks where rollout reward variance is highest under GRPO, preventing rapid depletion of informative samples. [59]
  • STARE uses surprisal-weighted token-level advantage reweighting to stabilize policy entropy during GRPO training, counteracting entropy collapse observed in RLVR pipelines. [60]
  • MAST performs mechanism-guided selective unlearning of RLVR-induced reasoning on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base with substantially lower collateral damage than full-parameter updates. [61]
  • Kolmogorov Regression lifts finite-dimensional diffusion policies to a Cameron-Martin function space to eliminate temporal drift from discretization in long-horizon robot deployment. [62]

VISION-LANGUAGE MODELS

  • ChartNet is a new MIT training dataset that teaches vision-language models to interpret charts accurately, targeting business analytics and scientific figure understanding. [63]
  • MIT Media Lab study finds that relying on AI for news fact-checking degrades users' own ability to detect misinformation, analogous to GPS weakening spatial navigation skills. [64]
  • ReSiReg addresses spatial inconsistency in dense VLM embeddings, which is critical for robots that must reason over semantics and 3D geometry simultaneously. [65]
  • Does VLA Know the Basics measures how much commonsense and world knowledge survives VLM-to-VLA fine-tuning, finding task failures often conflate missing knowledge with policy errors. [66]

MULTI-AGENT SYSTEMS

  • FoMoE breaks the full-replica barrier in Mixture-of-Experts pre-training by federating expert shards across loosely coupled nodes, reducing infrastructure coupling requirements. [67]
  • Leadership as Coordination Control maps when process-level coordination adds measurable value in multi-agent LLM teams, identifying a recovery-advantage boundary analogous to team-science theory. [68]
  • ARIADNE routes inference-time queries to the most appropriate task-specialized LoRA adapter without requiring task labels, enabling large adapter ecosystems on a shared backbone. [69]

CONTINUAL AND FEDERATED LEARNING

  • Catastrophic Forgetting is Low-Rank provides a function-space proof in the NTK regime that old-task prediction drift under new training is confined to low-rank output-space directions, enabling targeted regularization. [70]
  • C2FL proposes clustered continual federated learning to handle both spatial distribution shift and temporal concept drift across privacy-sensitive edge nodes. [71]

GRAPH NEURAL NETWORKS AND SCIENTIFIC AI

  • AGDN solves Traveling Salesman Problems with an Anisotropic Graph Diffusion Network that exploits directed edge structure more effectively than isotropic graph models. [72]
  • GNN-accelerated algebraic multigrid uses a data-driven smoother to break the dependence of pressure-Poisson solvers on mesh regularity in unstructured flow simulation. [73]
  • MIT Connor Coley lab builds AI models that embed chemical principles for drug compound discovery and design at the chemistry-ML interface. [74]

SECURITY AND GOVERNANCE OF AI SYSTEMS

  • OpenAnt combines LLM-driven code decomposition, adversarial verification, and dynamic testing to automate vulnerability discovery in large codebases beyond what static analysis or fuzzing alone achieves. [75]
  • Detecting Hidden ML Training With Zero-Overhead Telemetry evaluates adversarial robustness of GPU workload classification using only NVML counters, relevant to AI compute governance enforcement. [76]
  • NIST published a mathematical proof extending Gödel's incompleteness logic to AI systems, formally supporting a shift to continuous-monitor-and-update security models rather than one-time certification. [77]

📐 STANDARDS & POLICY

AI AGENT STANDARDS

  • NIST launched the AI Agent Standards Initiative in February 2026 to ensure the next generation of AI agents can function securely on behalf of users and interoperate across the digital ecosystem. [78]
  • NIST expanded its AI Consortium scope in May 2026, calling for new members and establishing six task groups focused on AI measurement science and evaluation across different application domains. [79]

AI SAFETY MEASUREMENT

  • NIST published a Gödel-inspired mathematical proof this week supporting a continuous-monitor-and-update security model for AI, arguing that static certification is formally insufficient for adaptive AI systems. [77]
  • A NIST-led team released a new AI model that identifies safe single-story fire evacuation routes step by step, with a multilevel version in development for complex building layouts. [80]

AUTONOMOUS VEHICLE LIABILITY

  • A new arXiv framework proposes hierarchical Bayesian credibility pricing for AV liability, pooling sparse incident data across cities, software versions, and territories to handle non-stationary risk under ODD shift. [81]

💰 FUNDING & PROGRAMS

NSF MAJOR INITIATIVES

  • NSF announced the 1.5 billion dollar X-Labs initiative on May 14, 2026, committing funds over the next decade toward generational breakthrough science efforts through novel collaborative structures. [82]
  • NSF relaunched SBIR and STTR programs on May 26, 2026 with 250 million dollars for startups and small businesses, including a new 40 million dollar pilot emphasizing next-generation scientific instrumentation. [83]
  • NSF renewed support for the MIT-led Institute for Artificial Intelligence and Fundamental Interactions, IAIFI, entering its second phase with increased funding at the AI-physics frontier. [84]

NSF AI EDUCATION

  • NSF-supported teams advanced through the inaugural Presidential AI Challenge, with a North Carolina State University-sponsored teacher named national champion on June 9, 2026. [85]
  • MIT RAISE and Georgia State University launched PATH on June 4, 2026, connecting universities, community colleges, industry, and government to expand industry-aligned AI training and career pathways. [86]
  • NSF-supported researcher Madhur Behl at the University of Virginia is using autonomous IndyCar racing to develop AI autonomy techniques and train the next-generation AI workforce. [87] [88]

DARPA PROGRAMS

  • DARPA published a physical intelligence research call on April 28, 2026, seeking to embed intelligence directly into robotic materials rather than rely solely on centralized computation. [89]
  • DARPA AI Forge released a new report and RFI on May 28, 2026, aiming to align government, academia, and industry around forward-looking national-security AI research priorities. [90]
  • DARPA invited the first wave of competitors to the Lift Challenge on June 8, 2026, with 6.5 million dollars in prizes at stake. [91]
  • DARPA celebrated 20 years of Young Faculty Awards on June 11, 2026, having supported more than 500 rising research stars from over 60 institutions, and announced new Director's Fellows. [92]

UKRI

  • BBSRC invested 10 million pounds in 21 new Fellows on June 18, 2026 as part of its commitment to develop the next generation of independent research leaders across the UK. [93]
  • UKRI backed Midlands Mindforge on June 10, 2026 to complete its first round of spinout investments supporting Midlands university startups. [94]
  • Innovate UK announced 22.9 million pounds across 18 projects in cycle five of the Ofgem Strategic Innovation Fund on June 9, 2026. [95]

📄 RESEARCH

WORLD MODEL SECURITY

  • Stealthy World Model Manipulation via Data Poisoning shows that adversarially poisoned fine-tuning trajectories can silently corrupt a model-based agent's world model, causing targeted planning failures without triggering anomaly detectors. In short, the training pipeline for world models is an under-examined attack surface. [96]

TACTILE SIM-TO-REAL

  • TactSpace learns a physics-enriched shared latent space that aligns simulated and real tactile sensor readings without requiring high-fidelity tactile simulation, which does not yet exist. The approach uses physics constraints during latent space training rather than physics simulation at inference time. [20]

RADAR PLACE RECOGNITION

  • Spatially Stratified Distillation enables a cheap 4D automotive radar to query reference maps built by an expensive spinning radar by distilling spinning-radar descriptors into a stratified spatial hierarchy the automotive sensor can match against. This is a practical step toward heterogeneous sensor interoperability in fleet mapping. [28]

INFERENCE-TIME POLICY IMPROVEMENT

  • VERITAS separates a robot's generative action policy from a visual verifier, using the verifier to score candidate action rollouts and steer the generator at test time. The system also uses verifier feedback to autonomously label new training data, enabling self-improvement without human annotation. [14]

ENERGY-BASED SURFACE CLASSIFICATION

  • A comparative analysis across three independent mobile-robotics datasets evaluates energy-derived features for terrain surface classification as both a standalone modality and as a supplement to established methods. Results suggest energy features are underexplored relative to their classification contribution, particularly under sensor constraints. [97]

That is your Robotics Pulse for June 19, 2026. See you tomorrow.

📎 Sources

  1. HALOMI: Learning Humanoid Loco-Manipulation with Active Perception from Human Demonstrations — arXiv cs.RO (Robotics)
  2. HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning — arXiv cs.RO (Robotics)
  3. Mobile Pedipulation for Object Sliding via Hierarchical Control on a Wheeled Bipedal Robot — arXiv cs.RO (Robotics)
  4. Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models — arXiv cs.RO (Robotics)
  5. ThinkingVLA: Interleaved Vision and Language Reasoning for Robotic Manipulation — arXiv cs.RO (Robotics)
  6. PearlVLA: Progressive Embodied Action-Plan Refinement in Latent Space — arXiv cs.RO (Robotics)
  7. MuseVLA: An Adaptive Multimodal Sensing Vision-Language-Action Model for Robotic Manipulation — arXiv cs.RO (Robotics)
  8. Motion-Focused Latent Action Enables Cross-Embodiment VLA Training from Human EgoVideos — arXiv cs.RO (Robotics)
  9. Object-Centric Residual RL for Zero-Shot Sim-to-Real VLA Enhancement — arXiv cs.RO (Robotics)
  10. WeaveLA: Event Driven Cross-Subtask Latent Memory Weaving for Repetitive Robot Manipulation — arXiv cs.RO (Robotics)
  11. GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planning — arXiv cs.RO (Robotics)
  12. WAM-RL: World-Action Model Reinforcement Learning with Reconstruction Rewards and Online Video SFT — arXiv cs.RO (Robotics)
  13. LAGO Policy: Latency-Aware Asynchronous Diffusion Policies with Goal-Directed Collision-Free Planning for Smooth Manipul — arXiv cs.RO (Robotics)
  14. Visual Verification Enables Inference-time Steering and Autonomous Policy Improvement — arXiv cs.RO (Robotics)
  15. DexLink Hand: A Compact, Affordable, 16-DOF Linkage-Driven Hand with Human-Like Dexterity — arXiv cs.RO (Robotics)
  16. Do as I Do: Dexterous Manipulation Data from Everyday Human Videos — arXiv cs.RO (Robotics)
  17. Zero-Shot Long-Horizon Dexterous Manipulation via Multi-View 3D-Grounded VLM Reasoning — arXiv cs.RO (Robotics)
  18. GCNGrasp-VP: Affordance-Guided View Planning for Efficient Task-Oriented Grasping — arXiv cs.RO (Robotics)
  19. EAGG: Embodiment-Aligned Grasp Generation via Geometry-Aware Graph Conditioning — arXiv cs.RO (Robotics)
  20. TactSpace: Learning a Physics-enriched Shared Latent Space for Tactile Sim-to-Real Transfer — arXiv cs.RO (Robotics)
  21. HT-Bench: Benchmarking and Learning Dexterous Full-Hand Tactile Representations with Egocentric Vision — arXiv cs.RO (Robotics)
  22. WireCraft: A Simulation Benchmark for Industrial DLO Manipulation — arXiv cs.RO (Robotics)
  23. FAST-LIVGO: A Degeneracy-Robust LiDAR-Inertial-Visual-GNSS Fusion Odometry — arXiv cs.RO (Robotics)
  24. RICH-SLAM: Radar SLAM with Incremental and Continuous Hilbert Mapping — arXiv cs.RO (Robotics)
  25. Sensor Configuration Matters: A Systematic Evaluation of Multimodal SLAM on Quadruped Robots — arXiv cs.RO (Robotics)
  26. C-ARC: Continuous-Adaptive Range Clustering for Non-Repetitive LiDAR Sensors — arXiv cs.RO (Robotics)
  27. A High-accuracy Event-based Underwater SLAM System — arXiv cs.RO (Robotics)
  28. Spatially Stratified Distillation for Heterogeneous Radar Place Recognition — arXiv cs.RO (Robotics)
  29. Viking Hill Dataset: A Lidar-Radar-Camera Dataset for Detection and Segmentation in Forest Scenes — arXiv cs.RO (Robotics)
  30. OneCanvas: 3D Scene Understanding via Panoramic Reprojection — arXiv cs.RO (Robotics)
  31. FLAP: FOV-Constrained Active Perception Planning for Prior-Map-Free 3D Navigation — arXiv cs.RO (Robotics)
  32. Monocular 3D Occupancy Perception for Robots on Sidewalks via Hybrid 2D-3D Learning — arXiv cs.RO (Robotics)
  33. ERQA-Plus: A Diagnostic Benchmark for Reasoning in Embodied AI — arXiv cs.RO (Robotics)
  34. ZiMPedance: Impedance-Aware ZMP Modeling and Control for Payload Carrying with Quadruped Robots — arXiv cs.RO (Robotics)
  35. Damage Adaptation in Seconds for Architected Materials — arXiv cs.RO (Robotics)
  36. Selective Unit-Cell Actuation in Lattice Structures for Distributed Morphology in Soft Robots — arXiv cs.RO (Robotics)
  37. Shape Sensing of Continuum Robots using Direct Laser Writing — arXiv cs.RO (Robotics)
  38. TerraTransfer: Learning End-to-End Driving Policies Without Expert Demonstrations — arXiv cs.RO (Robotics)
  39. Learning to Annotate Delayed and False AEB Events: A Practical System for Extreme Class Imbalance and Asymmetric Label N — arXiv cs.RO (Robotics)
  40. A Mixed-Reality Testbed for Autonomous Vehicles — arXiv cs.RO (Robotics)
  41. CABLE: Cloud-Assisted Bandwidth-efficient LMM-based Encoding for V2X Systems — arXiv cs.RO (Robotics)
  42. ED3R: Energy-Aware Distributed Disaster Detection Enabled by Cooperative Robotic Agents — arXiv cs.RO (Robotics)
  43. Accountability in Autonomous Drone-Based Firefighting: Insights From a Field Trial — arXiv cs.RO (Robotics)
  44. From Ad Hoc Pilots to Repeatable Patterns: Structuring Drone Collaboration in Emergency Services with DroneLets — arXiv cs.RO (Robotics)
  45. Generating Natural and Expressive Robot Gestures through Iterative Reinforcement Learning with Human Feedback using LLMs — arXiv cs.RO (Robotics)
  46. Seeing Through Occlusion: Deterministic Arm Kinematic Correction for Robot Teleoperation — arXiv cs.RO (Robotics)
  47. Beyond Failure Recovery: An Engagement-Aware Human-in-the-loop Framework for Robotic Systems — arXiv cs.RO (Robotics)
  48. Continual Online Personalization of Exoskeleton Control via Manifold-Aware Experience Replay — arXiv cs.RO (Robotics)
  49. Robotic Servicing of Geosynchronous Satellites technology to launch in 2026 — DARPA News
  50. MagicSim: A Unified Infrastructure for Executable Embodied Interaction — arXiv cs.RO (Robotics)
  51. GASE: Gaussian Splatting-Based Automated System for Reconstructing Embodied-Simulation Environments — arXiv cs.RO (Robotics)
  52. AnnotateAnything: Automatic Annotation of 3D Assets for Robot Manipulation — arXiv cs.RO (Robotics)
  53. EBench: Elemental Diagnosis of Generalist Mobile Manipulation Policies — arXiv cs.RO (Robotics)
  54. Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System — arXiv cs.RO (Robotics)
  55. Teaching AI agents to ask better questions by playing “Battleship” — MIT News — AI
  56. Diffusion-Proof: Recipe for Formal Theorem Proving Beyond Auto-Regressive Generation — arXiv cs.LG (Machine Learning)
  57. Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers — arXiv cs.AI (AI)
  58. Looped World Models — arXiv cs.AI (AI)
  59. RODS: Reward-Driven Online Data Synthesis for Multi-Turn Tool-Use Agents — arXiv cs.AI (AI)
  60. STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability — arXiv cs.AI (AI)
  61. Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoning — arXiv cs.AI (AI)
  62. Kolmogorov Regression for Robust Diffusion Policies — arXiv cs.AI (AI)
  63. MIT researchers teach AI models to interpret charts — MIT News — AI
  64. The consequences of relying on AI for accurate news — MIT News — AI
  65. ReSiReg: Towards Spatially Consistent Semantics in Language-Conditioned Robotic Tasks — arXiv cs.RO (Robotics)
  66. Does VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in Vision-Language-Action Models — arXiv cs.RO (Robotics)
  67. FoMoE: Breaking the Full-Replica Barrier with a Federation of MoEs — arXiv cs.AI (AI)
  68. Leadership as Coordination Control: Behavioral Signatures and the Recovery-Advantage Boundary in Multi-Agent LLM Teams — arXiv cs.AI (AI)
  69. ARIADNE: Agnostic Routing for Inference-time Adapter DyNamic sElection — arXiv cs.AI (AI)
  70. Catastrophic Forgetting is Low-Rank: A Function-Space Theory for Continual Adaptation — arXiv cs.LG (Machine Learning)
  71. C2FL: Clustered Continual Federated Learning under Spatial and Temporal Drift — arXiv cs.LG (Machine Learning)
  72. AGDN: Learning to Solve Traveling Salesman Problem with Anisotropic Graph Diffusion Network — arXiv cs.LG (Machine Learning)
  73. Acceleration of an algebraic multigrid pressure solver using graph neural networks — arXiv cs.LG (Machine Learning)
  74. Building AI models that understand chemical principles — MIT News — AI
  75. OpenAnt: LLM-Powered Vulnerability Discovery Through Code Decomposition, Adversarial Verification, and Dynamic Testing — arXiv cs.LG (Machine Learning)
  76. Detecting Hidden ML Training With Zero-Overhead Telemetry — arXiv cs.LG (Machine Learning)
  77. NIST Mathematical Proof Supports Transition to a Continuous-Monitor-and-Update Security Model for AI Systems — NIST News
  78. Announcing the "AI Agent Standards Initiative" for Interoperable and Secure Innovation — NIST News
  79. NIST Expands AI Consortium’s Scope, Calls for New Members — NIST News
  80. New AI Model Shows How to Evacuate for Fires One Safe Step at a Time — NIST News
  81. Credibility-Weighted Pricing of Autonomous Vehicle Liability Under Operational Design Domain Shift — arXiv cs.RO (Robotics)
  82. NSF announces $1.5B NSF X-Labs initiative to pursue generational breakthrough science efforts — NSF News
  83. NSF deploys $250 million to restart Small Business Innovation Research & Small Business Tech Transfer programs, includin — NSF News
  84. NSF renews support for MIT-led AI and physics institute, expanding a new model for discovery — MIT News — AI
  85. NSF-supported teams advanced through the Presidential AI Challenge, with North Carolina teacher named national champion — NSF News
  86. PATH to boost AI training and career opportunities for industry-aligned jobs — MIT News — AI
  87. Podcast: Autonomous IndyCar racing (part 1) — NSF News
  88. Podcast: Onboard autonomous IndyCar racing (part 2) — NSF News
  89. Rethinking robotics with physical intelligence — DARPA News
  90. AI Forge: Accelerating AI breakthroughs for national security — DARPA News
  91. DARPA invites first wave of Lift Challenge competitors — DARPA News
  92. DARPA celebrates 20 years of Young Faculty Awards, announces Director’s Fellows — DARPA News
  93. BBSRC invests in 21 new Fellows — UKRI News
  94. UKRI backs Midlands Mindforge to deploy first spinout investments — UKRI News
  95. Ofgem SIF: £22.9m to support ambitious, transformational projects — UKRI News
  96. Stealthy World Model Manipulation via Data Poisoning — arXiv cs.RO (Robotics)
  97. Leveraging Energy Features for Surface Classification with Deep Learning: A Comparative Analysis Across Three Independen — arXiv cs.RO (Robotics)

Curated from official sources — DARPA/NSF/NIST/IEEE/ORNL/MIT/UKRI/arXiv. Informational only.
Serial 20260619-00-v4 · 2026-06-19 00:02 UTC

Read next