Sign in Subscribe
Robotics

🤖 Robotics Pulse · 2026-07-24 00:01 UTC

ROBOTICS PULSE

Thursday, July 24, 2026

⚡ TL;DR

DARPA's Mission Robotic Vehicle has lifted off for geosynchronous orbit, marking the first operational robotic satellite servicing mission in history. Today's edition is dense with humanoid control breakthroughs, rapid robot skill acquisition, and a wave of NSF AI infrastructure funding totaling over $83M.

🤖 ROBOTICS

DARPA RSGS MISSION LAUNCHES

  • DARPA's Robotic Servicing of Geosynchronous Satellites (RSGS) Mission Robotic Vehicle is now en route to GEO, described by DARPA as a historic transformation of on-orbit servicing capability. [1]

DARPA AI-CONTROLLED F-16

  • In a separate DARPA milestone, the VENOM program flew an AI-controlled F-16 with the U.S. Air Force, demonstrating scalable AI development for operational fleet aircraft. [2]

HUMANOID MOTION TRACKING

  • The paper "What Matters in Humanoid General Motion Tracking" conducts a systematic empirical study of design choices for whole-body reference-following policies, isolating which factors actually drive balance and tracking performance. [3]
  • Extreme-RGMT introduces a continual learning framework that lets humanoid controllers progressively add highly dynamic skills such as acrobatics without eroding everyday locomotion abilities. [4]

RAPID SKILL ACQUISITION FROM VIDEO

  • HOST (Human-to-robot One-Shot Transfer) lets robots acquire manipulation skills in seconds from a single human video at inference time, with no retraining loop, avoiding the catastrophic forgetting that plagues current methods. [5]

ROBOT GUIDE DOG FOR BLIND USERS

  • Milo is presented as a fully autonomous indoor/outdoor robotic guide dog targeting Blind and Low-Vision users, addressing the roughly $50,000 cost and long wait lists of trained biological guide dogs. [6]

EGOCENTRIC FAILURE RECOVERY

  • EgoRecovery trains robot policies to recover from manipulation failures using human recovery demonstrations captured from an egocentric viewpoint, providing recovery behavior data without expensive robot teleoperation. [7]

MIT SCENESMITH ROBOT TRAINING DATA

  • MIT's SceneSmith system uses collaborative AI agents to procedurally generate realistic 3D kitchen, hotel, and living-room environments, giving robots synthetic training data for everyday chores without manual scene authoring. [8]

RETAIL HUMANOID VLA FRAMEWORK

  • DEED (Data-Efficient Post-Training and Experience-Driven Learning) is a VLA framework specifically targeting the lab-to-store gap for humanoid robots, addressing execution errors, distribution shift, and environmental variability in retail settings. [9]

MODPACK TELEOPERATION SYSTEM

  • ModPack is a modular, extensible teleoperation interface supporting diverse robot embodiments and bimanual mobile manipulation within a single unified framework, directly addressing hardware lock-in in data collection pipelines. [10]

LENS: LLM-GUIDED CLUTTER MANIPULATION

  • LENS uses an LLM to simplify cluttered scenes for robotic planning and control, reducing the combinatorial complexity from multiple objects, collisions, and task ambiguity that defeats conventional manipulation stacks.

AUTONOMOUS DRIFTING FOR COLLISION AVOIDANCE

  • Researchers demonstrate emergent autonomous drifting for collision avoidance in real-world winter driving scenarios, moving beyond engineered test tracks to validate high-sideslip maneuvers in naturalistic conditions.

MIT CHIP FOR TINY ROBOT NAVIGATION

  • MIT researchers combined an efficient algorithm with dedicated hardware to generate 3D navigation maps on minimal memory and power, directly targeting tiny robots that cannot carry conventional compute loads.

SOFT ROBOT ACOUSTIC VICTIM LOCALIZATION

  • A distributed acoustic localization array deployed along the body of a soft everting vine robot can find disaster victims in confined, unstructured environments using sound-based sensing.

SUBSEA AUV PIPELINE INSPECTION OPTIMIZATION

  • A two-stage mixed-integer linear programming framework jointly optimizes Subsea Docking Plate placement and resident AUV allocation to minimize maximum and average pipeline incident response times.

SEEDEDGRASP: LANGUAGE-GUIDED GRASPING

  • SeededGrasp combines VLM language grounding with 3D spatial reasoning to select grasps across multiple robot embodiments in complex, cluttered scenes without requiring embodiment-specific retraining.

SURGICAL EVERSION ROBOT

  • Eversion-based robots are shown to enable safe access, steering, and endoscopic imaging within the narrow and compliant spinal subarachnoid space, avoiding the friction and shear that disable conventional catheters.

MVP-TAC SURGICAL TACTILE SENSOR

  • MVP-Tac is a miniaturized dual-modal vision and photoelastic tactile sensor for robot-assisted minimally invasive surgery, addressing the longstanding absence of palpation feedback in laparoscopic robotic systems.

🧠 AI & MODELS

SLAI T-REX: TRILLION-PARAMETER POST-TRAINING

  • SLAI T-Rex describes full-parameter post-training of the DeepSeek-V4 family at trillion-parameter MoE scale on Ascend SuperPOD, tackling memory pressure, communication overhead, and kernel inefficiency that block this regime.

OFF-CONTEXT GRPO: REASONING ON HARD PROBLEMS

  • Off-Context GRPO addresses the zero-learning-signal failure of RLVR on hard problems by injecting privileged information during training, enabling models to learn from problems they cannot yet solve unaided.

PRO-LONG: PROGRAMMATIC MEMORY FOR LONG-HORIZON REASONING

  • PRO-Long introduces programmatic memory to LLM agents, targeting persistent failures on long-horizon continual learning benchmarks such as ARC-AGI-3 where standard context windows fall short.

DYNAMICRUBRIC: CO-EVOLVING EVALUATORS AND POLICIES

  • DynamicRubric co-evolves LLM evaluators alongside the policies they assess, preventing the collapsed relative scoring that stalls policy optimization when response quality becomes nearly uniform.

ISO: RLVR OPTIMIZATION STACK

  • ISO is an RLVR-native optimization stack that studies the poorly understood gap between reward feedback and weight-space updates in reinforcement learning with verifiable rewards for language model reasoning.

RESEARCHARENA: AI SABOTAGE MONITORING

  • ResearchArena is a benchmark for evaluating covert sabotage and monitoring in automated AI R&D pipelines, treating the AI agent as a potential adversary and testing whether monitors can detect misaligned behavior.

SOUND PROBABILISTIC SAFETY BOUNDS FOR LLMS

  • Researchers apply Clopper-Pearson confidence intervals to derive probably-approximately-correct (PAC) bounds on the probability that an LLM generates harmful output to a given prompt, providing the first rigorous safety certificates of this kind.

ABOT-WORLD-0: REAL-TIME WORLD MODEL ON DESKTOP GPU

  • ABot-World-0 is an action-conditioned video world model that runs long-horizon closed-loop interaction in real time on a single desktop GPU, trained on AAA games, simulation engines, and internet video.

KOOPMAN DREAMER: STABLE LATENT WORLD MODELS

  • Koopman Dreamer imposes spectral constraints on latent world model transitions, directly controlling modal persistence and limiting error accumulation during long imagination rollouts used for policy optimization.

DIFFUSION REROLL: REVISABLE ROBOT PREDICTION

  • Diffusion ReRoll enables revisable denoising for robotic sequential prediction by selectively re-noising problematic regions of a predicted horizon, unlike single-pass diffusion predictors that cannot correct partial errors.

HOST SKILL TRANSFER (see also ROBOTICS)

  • HOST is grounded in a one-shot inference-time paradigm: a single human video supplies the skill template, and the robot adapts without any retraining, preserving all previously mastered skills. [5]

GEMMA-4 MATERIALS SCIENCE REPRESENTATIONS

  • Researchers find that materials science mechanism information in Google's open-weight Gemma-4-E4B-it model has three experimentally separable representational forms, showing the model encodes governing physics rather than producing correct outputs by surface pattern matching.

AGENTBENCH: WHERE RESEARCH MEETS DEPLOYMENT

  • A survey paper maps the transition of LLM-based agentic systems from research prototypes to production deployments across software engineering, scientific discovery, and multi-agent coordination, identifying open reliability and safety gaps.

LICENSE LAUNDERING IN AI SUPPLY CHAINS

  • A study of the Hugging Face and GitHub AI artifact supply chain finds that license obligations frequently fail to propagate through redistribution, creating legal exposure that neither dataset nor model cards currently flag.

GENERATIVE AI FLOODING THE BOOK MARKET

  • Full-text AI detection across 14,419 self-published genre fiction books finds measurable market dilution from near-zero-cost AI-generated content, challenging the assumption that buyers will naturally filter out low-quality AI output.

📐 STANDARDS & POLICY

IEEE AI ETHICS CERTIFICATION

  • IEEE SA's ICAP program offers a Certified AI Ethics Professional (CAEGP) credential, providing structured team-level skills for responsible AI deployment, with guidance published July 20, 2026 on post-certification career use.

IEEE AT GENEVA DIGITAL WEEK 2026

  • IEEE participated in Geneva Digital Week (July 6-10, 2026), engaging governments, international organizations, and technical communities on global digital governance frameworks directly relevant to AI and robotics standards.

NIST NEW DIRECTOR CONFIRMED

  • Arvind Raman, formerly dean of engineering at Purdue University, has been confirmed as the 18th NIST Director, taking the helm of the agency that leads U.S. AI measurement and standards work.

NIST MEP FUNDING FOR MANUFACTURERS

  • NIST announced a funding opportunity for 14 Manufacturing Extension Partnership centers to advance small and medium-sized U.S. manufacturers, with an informational webinar scheduled July 28, 2026.

AI SAFETY FAILURES WE ARE NOT INSTRUMENTING

  • An arXiv perspective paper argues current AI safety discourse over-indexes on visible, dramatic failures and misses quieter deployed-system failures that are plausible, hard to detect, and consequential, calling for new instrumentation.

💰 FUNDING & PROGRAMS

NSF GENESIS MISSION AND AI DATA INFRASTRUCTURE

  • NSF joined the White House OSTP and DOE in the Genesis Mission to advance administration AI priorities, with MIT projects selected covering natural resources, manufacturing, and nuclear physics.
  • NSF announced an $83 million investment through the Integrated Data Systems and Services (IDSS) program to expand data infrastructure for AI-driven science and U.S. research competitiveness.
  • NSF launched the Unlocking Dataset Value for AI-Enabled Scientific Discovery program, targeting scientific community datasets specifically to enable AI-based research discovery.

UKRI FIVE-YEAR AI AND QUANTUM ROADMAP

  • UKRI published a five-year strategy to back more than 20,000 researchers in AI and quantum, targeting breakthrough discoveries and national growth through an ambitious technology roadmap.

UKRI LAUNCHES TWO AI RESEARCH LABS

  • EPSRC under UKRI launched two new AI research labs to develop next-generation AI systems, positioning the UK to compete in the global AI race.

NSF REGIONAL INNOVATION ENGINES

  • NSF awarded 12 new Regional Innovation Engines across 20 U.S. states to build and scale innovation clusters accelerating research and economic growth in manufacturing, AI, and related sectors.

📄 RESEARCH

NAVVERSE INDOOR-TO-OUTDOOR NAVIGATION BENCHMARK

  • NavVerse is a new continuous-simulation benchmark requiring robots to navigate seamlessly from building interiors to outdoor streets in a single episode, filling a critical gap left by benchmarks that treat indoor and outdoor navigation as separate problems. Delivery, campus, and emergency-response deployments are the target applications.

KINEBENCH: EVALUATING EMBODIED WORLD MODELS

  • KineBench evaluates the physical consistency of embodied world models using IDM-Free kinematic grounding, avoiding dependence on Inverse Dynamics Models that can mask physically inconsistent predictions, enabling more faithful closed-loop assessment.

DINS-IO: LABEL-FREE INERTIAL ODOMETRY

  • DINS-IO trains inertial odometry directly from raw IMU streams without expensive motion-capture position ground truth, using differentiable INS consistency as the self-supervised training signal, which could dramatically lower the cost of deploying learned odometry on field robots.

EA-NAV: EMBODIMENT-AWARE VISUAL NAVIGATION

  • EA-Nav addresses cross-embodiment navigation by encoding robot-specific physical constraints alongside visual observations, preventing the prediction ambiguity that arises when the same visual scene implies different safe actions for differently sized or shaped agents.

AGENTIC REAL2SIM: PHYSICS-BASED WORLD MODELING

  • Agentic Real2Sim uses vision-language agents to automate real-to-simulation conversion for robotic object interaction, recovering scene geometries, physical parameters, and actor configurations from real-world observations with minimal human labor.

SOPD-SOCIALNAV: LIGHTWEIGHT SOCIAL NAVIGATION VIA DISTILLATION

  • SOPD-SocialNav uses selective on-policy distillation to compress large vision-language model social navigation capabilities into lightweight models deployable on resource-constrained robots, preserving semantic scene understanding without requiring cloud inference.

In Memoriam: Professor Emeritus Dimitri Bertsekas of MIT, whose work on control, optimization, and large-scale computation shaped the mathematical foundations of modern robotics and AI, died July 22, 2026, at age 83.

ROBOTICS PULSE is compiled from official sources including DARPA, NSF, NIST, IEEE SA, UKRI, MIT News, ORNL, and arXiv cs.RO/cs.AI/cs.LG. All claims are grounded in cited items.

📎 Sources

  1. Robotic Servicing of Geosynchronous Satellites lifts off — DARPA News
  2. DARPA, U.S. Air Force fly AI-controlled F-16 — DARPA News
  3. What Matters in Humanoid General Motion Tracking? An Empirical… — arXiv cs.RO (Robotics)
  4. Extreme-RGMT: Continual Learning of Highly Dynamic Skills for … — arXiv cs.RO (Robotics)
  5. Robots Acquire Manipulation Skills in Seconds from a Single Hu… — arXiv cs.RO (Robotics)
  6. Milo, a Fully Autonomous Indoor/Outdoor Robotic Guide Dog — arXiv cs.RO (Robotics)
  7. EgoRecovery: Acquiring Failure Recovery Ability Through Human … — arXiv cs.RO (Robotics)
  8. AI agents create virtual playgrounds to help robots get crucia… — MIT News — AI
  9. Closing the Lab-to-Store Gap: A Data-Efficient Post-Training a… — arXiv cs.RO (Robotics)
  10. ModPack: An Extensible Teleoperation Interface for Bimanual Mo… — arXiv cs.RO (Robotics)

Curated from official sources — DARPA/NSF/NIST/IEEE/ORNL/MIT/UKRI/arXiv. Informational only.
Serial 20260724-00-v39 · 2026-07-24 00:01 UTC · pulse.uzylab.com

Read next