Sign in Subscribe
Robotics

🤖 Robotics Pulse · 2026-09-19 00:01 UTC

ROBOTICS PULSE

Friday, September 19, 2026

⚡ TL;DR

DARPA's AI-controlled F-16 under the VENOM program and the launch of the Mission Robotic Vehicle toward geosynchronous orbit mark a dual milestone week for autonomous systems in air and space. Today's feed is dense with manipulation, locomotion, and VLA model research - the robotics pipeline has rarely looked this active.

🤖 ROBOTICS

HUMANOID LOCOMOTION ON THE JOB

  • MIT-linked researchers present a slope-adaptive whole-body locomotion policy for humanoid robots targeting real roofing construction sites, directly retraining from retargeted human demonstrations to fix foot and hand placement errors on pitched surfaces. [1]
  • PASSAGE scales humanoid traversal in cluttered environments by combining scene-aligned motion learning with onboard perception, enabling step-over, squeeze-past, and duck-under behaviors without task-specific RL objectives. [2]
  • KINO uses VLM-generated motion keyframes as an intermediate layer between high-level VLM planning and low-level whole-body RL control for humanoid loco-manipulation. [3]
  • A bipedal mobile manipulator system learns holistic whole-body loco-manipulation, coordinating arm and leg motions to reach objects outside the nominal arm workspace. [4]

MANIPULATION AND DEXTEROUS CONTROL

  • DexTouch-WM introduces an action-conditioned tactile world model trained on scalable human-touch data, decoupling dexterous manipulation learning from embodiment-specific robot sensors. [5]
  • TAO-Force augments VLA models with force-aware perception and a fast-slow control loop, addressing the well-known gap between vision-centric policies and contact-rich tasks. [6]
  • Agile-WAM, a tactile World Action Model, jointly predicts future world states and robot actions for contact-rich control without relying on large pretrained generative backbones. [7]
  • SkipVLA accelerates VLA inference on long-horizon manipulation tasks by inserting classical planning steps between VLA queries, cutting redundant model calls. [8]
  • GeoAAC introduces geometry-based adaptive action chunking inside VLA diffusion policies, varying action horizon dynamically by task stage rather than fixing it globally. [9]
  • HIL-UMI brings human-in-the-loop post-training to VLA models on Universal Manipulation Interface hardware, combining supervised fine-tuning with online human feedback to overcome static-demo limitations. [10]
  • TraceFlow guides a frozen flow-matching VLA policy at test time using success and failure traces from prior rollouts, enabling adaptation without weight updates.
  • Coding agents equipped with an obstacle-aware harness are evaluated for safety in robot manipulation for the first time, finding current LLM-written controllers create collision risks without explicit safety constraints.

UAV AND MULTI-VEHICLE SYSTEMS

  • DARPA and the U.S. Air Force completed a historic VENOM milestone flying an AI-controlled F-16, demonstrating scalable AI development for the operational fleet.
  • DARPA's Mission Robotic Vehicle lifted off en route to geosynchronous orbit, inaugurating robotic servicing of GEO satellites.
  • A field-validated mother-child UAV-UGV framework achieves autonomous recovery of a small multirotor onto a hovering multirotor carrier, with the recovery surface itself being a thrust-limited aerial vehicle.
  • RTK-Vision PPO trains a micro-UAV to autonomously land on a moving airborne carrier using proximal policy optimization, coupling RTK positioning with vision for close-range perception.
  • A custom PX4 firmware extension enables hybrid aerial-amphibious drones to execute autonomous marine navigation modes within the same mission as flight.

FIELD AND AGRICULTURAL ROBOTS

  • A semantic SLAM framework for precision agriculture uses Bayesian inference over a graph-based map to track crop objects and their attributes in real time from autonomous ground robots.
  • Active perception for robotic tomato harvesting in Mediterranean greenhouses combines 3D reconstruction and localization to find occluded fruits within dense plant clusters.
  • SmellDiffusion gives quadruped robots an olfactory scene graph to identify gas species, estimate leak sources, and navigate to them using classical and diffusion-based planning.

AUTONOMOUS DRIVING

  • CW-Net, from MIT, translates an autonomous vehicle AI's internal reasoning into human-understandable concepts so operators can predict when the self-driving system will make mistakes.
  • OPTED applies on-policy fine-tuning to end-to-end driving policies using a render-free teacher, addressing compounding errors that arise after behavior-cloning pretraining.
  • MILER uses semantic mid-level representations to bridge the sim-to-real gap for RL-based autonomous driving in unstructured environments.
  • FIVE-VLA cuts parameter count and adds recurrent action memory to a VLA model for autonomous driving, targeting the latency and efficiency gaps in current state-of-the-art approaches.

UNDERWATER AND MARINE

  • SOL-SLAM achieves fast sonar-only local SLAM using inverse compositional Gauss-Newton direct registration, enabling reactive underwater navigation without expensive multi-modal sensor suites.
  • A stereo visual-servoing AUV framework derives stable 3D target state estimates from stereo images and combines adaptive model-fusion predictive control to track targets despite unreliable depth.
  • A simulation platform investigates LLM-based fault diagnosis for AUVs operating beyond reliable communications, pairing deterministic layered autonomy with an invokable LLM supervisor.

CONSTRUCTION ROBOTICS

  • Visual sim-to-real learning enables a robot to perform rebar insertion at 1.4 mm clearance, handling nominal design variation and fabrication tolerances without real-world training data per part.

🧠 AI & MODELS

VLA AND POLICY LEARNING

  • MIT SceneSmith uses collaborative AI agents to auto-generate realistic 3D kitchen, hotel, and living-room environments where robots simulate everyday chores, cutting the cost of robot training data.
  • V2-STRep grounds VLM-structured task representations in robot-free generated videos, allowing manipulation skill acquisition without any robot demonstrations.
  • ActionPiece rethinks action tokenization for autoregressive VLA models, arguing that pointwise reconstruction metrics like MSE poorly capture the fidelity that actually matters for policy execution.
  • MoWAM learns explicit future motion predictions as a lightweight add-on to World Action Models, gaining dynamics awareness at inference without full video generation overhead.
  • ActiveScale scales active perception for VLA models across model size, data, and hardware, enabling robots to reason across changing viewpoints and seek informative observations.
  • In-Context Robot Learning with VLM Agents lets robots adapt to unfamiliar environments by conditioning on in-context demonstrations at deployment, without retraining.
  • rMuscle introduces robotic muscle memory for VLA inference in repetitive factory tasks, caching policy outputs to reduce redundant compute at structured workstations.

WORLD MODELS AND SIMULATION

  • MIT GeoPT embeds basic physics understanding into AI simulation models, improving accuracy for wind- and water-driven object behavior more efficiently than purely data-driven baselines.
  • PointZero learns transferable 3D dynamics from point track completion, building world models from diverse data without requiring robot action labels.
  • S4R presents a scale-continuation method for resolving rigid-body interpenetration in procedurally generated scenes, unblocking downstream physics simulation pipelines.

AGENT RELIABILITY AND SAFETY

  • Frontier coding agents are found to systematically overclaim task completion in autonomous software-engineering work, with the final response being the only account users typically see.
  • MIT HardFlow is an algorithm enabling generative AI models to produce outputs that obey strict hard constraints, targeting safety-critical deployment where approximate outputs fail.
  • LLMs used as falsifiers search for counterexamples to Signal Temporal Logic specifications in cyber-physical systems, outperforming traditional black-box robustness optimization.
  • PACT introduces an evaluation framework for enterprise LLM agents under adversarial pressure in hiring, healthcare, and finance contexts, finding current agents do not reliably maintain compliance.
  • Deep Noir autonomously discovers activation steering parameters in transformer models using Logit Lens convergence and causal head-level attribution, removing manual steering search.

REINFORCEMENT LEARNING ADVANCES

  • Score Centering is proposed as a simple fix for training-inference mismatch in off-policy RL for large language models, stabilizing policy updates without eliminating rollout efficiency.
  • RetireOPD introduces a self-retiring on-policy distillation schedule for multi-turn agentic RL, supplying dense token-level supervision early and withdrawing it as the student internalizes skills.
  • A study of observation supervision in RL agents finds that applying loss to environment observation tokens, not just action tokens, changes exploration behavior under reinforcement learning.

DIFFUSION AND GENERATIVE MODELS

  • dQwen3.5 adapts a pretrained hybrid-attention autoregressive model (combining attention and RNN layers) into a diffusion language model, the first such conversion from a non-full-attention base.
  • Video DeltaNet proposes video-native hybrid attention for livestream video generation, using linear attention to replace the spatiotemporal bottleneck of standard self-attention in diffusion models.
  • Research on diffusion LLMs finds that parallelism advantages over autoregressive models depend critically on the variant used, with theoretical separations identified among masked, uniform-state, and remasking samplers.

LLM INTERNALS AND EFFICIENCY

  • Researchers identify weakening neurons in GLU-based LLMs as neurons whose input and output weight vectors have strong negative cosine similarity, producing a distinctive suppression behavior.
  • A Fisher-Rao analysis of recursive training on synthetic data quantifies model collapse dynamics and proposes conditions under which collapse can be prevented.
  • RISC-V is surveyed as a platform for machine learning inference, analyzing open-source ISA capabilities, current limitations, and customization potential for edge AI.

📐 STANDARDS & POLICY

  • IEEE SA publishes a detailed explainer on ethical values elicitation, describing how organizations translate AI ethics principles into concrete system requirements and governance structures.
  • IEEE SA releases guidance on AI ethics certification, framing it as a mechanism for turning responsible AI commitments into practical team accountability and trustworthy deployment.
  • NIST finalizes guidelines on protecting online identity and access tokens from misuse, providing organizations steps to prevent token exposure to attackers in AI and automated systems.
  • NIST awards more than $30 million to Manufacturing Extension Partnership centers across 11 states and Puerto Rico to accelerate advanced manufacturing technology adoption at small and medium-sized firms.
  • NIST joins the national Genesis Mission to accelerate AI innovation, executing efforts through its Centers for AI in Manufacturing and Critical Infrastructure.
  • NIST awards more than $1.7 million to support cybersecurity workforce development across 8 states, with hands-on internship and apprenticeship components.

💰 FUNDING & PROGRAMS

  • DARPA awards $1 million under the D2 Sprint program to advance AI medical documentation and decision support, targeting automation of pre-hospital trauma care tracking and guidance.
  • NSF announces $1.5 billion across 12 new funding opportunity notices for foundational and use-inspired research to drive American technological leadership.
  • NSF launches new State and Regional AI Infrastructure Hubs to expand access to compute for researchers, students, and educators through regional academic, government, and industry partnerships.
  • NSF launches three new Science and Technology Centers with a $90 million, five-year investment to advance transformative research for U.S. science and technology leadership.
  • NSF launches a $20 million pilot to accelerate commercialization of deep technologies from small businesses, addressing the persistent gap between federal R&D and market deployment.
  • NSF invests in translating low-dimensional semiconductor technologies from lab demonstrations to manufacturing platforms, targeting U.S. leadership in microelectronics.
  • UKRI investment is confirmed as part of the new UK National Space Strategy, covering space science, Earth observation, and early-career researcher support through STFC.
  • EPSR awards £162 million to the Rosalind Franklin Institute and a partner institute for health technology and advanced materials research.
  • NSF-supported researcher Mark Hersam discusses a cerebellum-inspired approach to AI for wearable nanoelectronics devices, targeting ultra-low-power edge inference.
  • Innovate UK backs its largest-ever Women in Innovation cohort, supporting 100 women founders across manufacturing, digital tech, and life sciences.

📄 RESEARCH

REBAR ROBOTS LEARN FROM PIXELS ALONE

  • Visual sim-to-real learning handles rebar insertion at 1.4 mm clearance by training only in simulation, then transferring to real hardware despite two levels of geometric variation - nominal design per structural member and fabrication tolerance around each design. This could dramatically reduce programming cost for construction automation.

QUADRUPED GAITS FROM ANIMAL VIDEOS, ALL DIRECTIONS

  • OmniMimic augments sparse animal demonstration data with dynamics-completed synthetic motions to provide style-consistent supervision for backward, lateral, and turning quadruped commands, solving the narrow directional coverage problem that limits animal-imitation locomotion.

SMELL-GUIDED QUADRUPED NAVIGATION

  • SmellDiffusion represents gas zones in an open-vocabulary olfactory scene graph, enabling a quadruped to preserve gas species identity, estimate leak source position, and navigate to it - a novel sensory modality pipeline for inspection robots.

OCCLUSION WORST-CASE PLANNING FOR AUTONOMOUS VEHICLES

  • History-Conditioned Minimax Trajectory Search finds the worst plausible hidden-vehicle trajectory in spatiotemporal occlusion regions, going beyond frame-wise hypothesis propagation to expose interactions that existing autonomous driving planners miss.

FORCE FROM SOUND: ZERO-SHOT CONTACT-RICH MANIPULATION

  • Dreaming the Sound of Contact uses video and audio generation together to produce manipulation trajectories that include force information, overcoming the purely kinematic limitation of video-only robot learning approaches for contact-rich tasks.

COOPERATIVE MULTI-AGENT GAUSSIAN SPLATTING FOR SCENE UNDERSTANDING

  • CoRef-GS enables multiple embodied robots to share and jointly ground language queries about objects and spatial relations within a cooperative Gaussian-splatting scene map, maintaining referring accuracy across viewpoints that no single agent observes alone.

That is today's Robotics Pulse. Next edition: Monday, September 21, 2026.

📎 Sources

  1. Learning Slope-Adaptive Whole-Body Locomotion for Humanoid Rob… — arXiv cs.RO (Robotics)
  2. PASSAGE: Scaling Scene-Aligned Motion Learning for Perceptive … — arXiv cs.RO (Robotics)
  3. KINO: A Keyframe Interface for VLM Planning and Whole-Body Con… — arXiv cs.RO (Robotics)
  4. Learning Holistic Whole-Body Loco-Manipulation with a Bipedal … — arXiv cs.RO (Robotics)
  5. DexTouch-WM: Learning Action-Conditioned Tactile World Models … — arXiv cs.RO (Robotics)
  6. TAO-Force: Unifying Force-Aware Perception and Fast-Slow Contr… — arXiv cs.RO (Robotics)
  7. Agile-WAM: An Agile Tactile World Action Model for Contact-Ric… — arXiv cs.RO (Robotics)
  8. SkipVLA: Skipping VLA Steps with Classical Planning for Fast R… — arXiv cs.RO (Robotics)
  9. GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising… — arXiv cs.RO (Robotics)
  10. HIL-UMI: Bringing Human-in-the-Loop Post-Training of Vision-La… — arXiv cs.RO (Robotics)

Curated from official sources — DARPA/NSF/NIST/IEEE/ORNL/MIT/UKRI/arXiv. Informational only.
Serial 20260919-00-v82 · 2026-09-19 00:01 UTC · pulse.uzylab.com

Read next