Sign in Subscribe
Robotics

🤖 Robotics Pulse · 2026-07-23 00:01 UTC

ROBOTICS PULSE

2026-07-23 | Your daily briefing on robotics and AI

⚡ TL;DR

DARPA and the U.S. Air Force completed a historic AI-controlled F-16 flight under the VENOM program, marking a scalable autonomy milestone for operational fleets. Today's edition is heavy on robotics hardware and manipulation research, with a major U.S. funding surge for AI-driven science infrastructure also landing this cycle. [1] [2]

🤖 ROBOTICS

AERIAL-TERRESTRIAL HYBRID LOCOMOTION

  • DASH (Ducted Aerial Spring Hopper) combines ducted-fan flight with spring-hopping ground locomotion in a mechanically minimalistic design using contact-implicit optimal control. [3]

TINY ROBOT NAVIGATION CHIP

  • MIT researchers paired a dedicated hardware chip with an efficient algorithm to generate 3D navigation maps using minimal memory and power, targeting micro-robot traversal of complex environments. [4]

VISION-LANGUAGE-ACTION PRECISION

  • STeP proposes using Signal Temporal Logic to encode spatial, temporal, and logical constraints for VLA models, giving robots formal, verifiable instruction compliance beyond natural language alone. [5]

EMBODIED WORLD MODELING SUITE

  • RoboInter1.5 releases a holistic intermediate representation dataset with fine-grained annotations for long-horizon manipulation, execution, and environment dynamics simulation, building on the prior RoboInter1.0 framework. [6]

HUMANOID LOCO-MANIPULATION

  • A new paper on Closing the Loop in Humanoid VLA introduces persistent 3D object tokens so humanoid robots track task objects across movement, contact, and occlusion during long-horizon loco-manipulation. [7]

FORCE-AWARE ROBOT MEMORY

  • FM-VLA adds force-based memory to vision-language-action models for contact-rich manipulation, going beyond vision-only or text-summary memory to capture haptic history. [8]

SURGICAL EVERSION ROBOT

  • Researchers demonstrate an eversion-based robot for navigating the narrow spinal subarachnoid space, eliminating the friction and shear generated by conventional catheter-pushing approaches. [9]

DUAL-MODAL SURGICAL TACTILE SENSOR

  • MVP-Tac is a miniaturized sensor combining vision and photoelastic tactile feedback for robot-assisted minimally invasive surgery without sacrificing visual navigation bandwidth. [10]

FABRIC PNEUMATIC ARTIFICIAL MUSCLES

  • A new drawstring-inspired fabric PAM design eliminates the excess radial deformation of conventional pneumatic muscles during axial contraction, saving working space in soft robotic actuators.

AUTONOMOUS DRILLING BOOM

  • The Sandvik SB60 long-reach drilling boom gets a collision-aware autonomy upgrade using pose-parameterized planning and CBF-QP self-collision filtering.

LEGGED BALANCE CONTROL

  • Koopman DCM extends divergent component of motion balance control beyond linear inverted pendulum models using data-driven Koopman unstable eigenfunctions for general legged systems.

GPS-FREE UAV POSITIONING

  • NGPS provides absolute positioning for high-altitude UAVs by matching downward-facing camera images to georeferenced satellite imagery with deep features and multi-rate sensor fusion.

UAV TEST-TIME SCALING

  • A paper shows that test-time scaling of VLMs, applying multiple inference passes without additional training, meaningfully improves UAV vision-language navigation in challenging scenes.

AI-CONTROLLED F-16 MILESTONE

  • DARPA and the U.S. Air Force flew an AI-controlled F-16 under the VENOM program, framing the achievement as demonstrating scalable AI development applicable to the broader operational fleet. [1]

DARPA LIFT CHALLENGE

  • Over 120 teams are competing for $6.5 million in DARPA Lift Challenge prizes to develop novel heavy-lift drone designs.

OPEN ANT ROBOT PLATFORM

  • The Open Ant is an open-hardware quadruped platform designed to close the sim-to-real gap for RL locomotion researchers, reducing reliance on simulation-only methodology.

REAL2SIM WITH VISION-LANGUAGE AGENTS

  • Agentic Real2Sim uses VLA agents to automate scene geometry recovery, physical parameter inference, and simulation assembly for robotic interaction, cutting the manual labor of real-to-sim pipelines.

MULTI-AGENT WAREHOUSE COORDINATION

  • A new Lifelong Multi-Subsystem Pickup and Delivery framework handles payload handovers between region-confined agents through shared buffer-limited stations, addressing a gap in lifelong MAPD.

🧠 AI & MODELS

WORLD MODEL FOR ROBOT ACTION

  • Masked Visual Actions (MVA) proposes representing robot actions as masked visual regions aligned with video model representations, letting pretrained video world models serve as robotic action substrates.

WORLD ACTION MODEL WITH REASONING MEMORY

  • WorldScape Policy 2.0 extends World Action Models with reasoning-augmented memory, long temporal context, and multi-modal conditioning to improve robotic manipulation planning.

SUBGOAL PLANNING UNDER UNCERTAINTY

  • S3 stabilizes subgoal selection in Hierarchical RL by constraining uncertainty of coarse dynamics, improving long-horizon task performance where flat RL fails.

OFFLINE RL WITH EXPERT QUERIES

  • A new paper combines conservative querying and adaptive regularization to improve offline RL under uncertainty, using expert feedback without additional environment interaction.

ADAPTIVE SPECULATIVE DECODING

  • AdaFlash uses on-policy distilled diffusion drafters to accelerate LLM inference via speculative decoding, building on the DFlash drafting paradigm with adaptive policy alignment.

RLVR OPTIMIZATION STACK

  • ISO (RLVR-Native Optimization Stack) systematically studies and improves the weight-update layer that converts verifiable reward signals into parameter changes for reasoning LLMs.

OFF-CONTEXT GRPO FOR HARD REASONING

  • Off-Context GRPO addresses the zero-signal problem in RLVR when models cannot generate any correct solutions, by introducing privileged guidance at training time for difficult problems.

LLM INFERENCE ON EDGE

  • SelectInfer selectively loads and computes only relevant neurons for on-device LLM inference, targeting resource-constrained edge hardware without model retraining.

AI POWER CONSUMPTION ESTIMATION

  • MIT's EnergAIzer method estimates AI power consumption in seconds, enabling data center operators to allocate resources efficiently and cut wasted energy.

AI-CONTROLLED F-16 AS AI CAPABILITY SIGNAL

  • The DARPA VENOM F-16 flight is explicitly framed as validating scalable AI autonomy development methodology, not just a one-off demo, with implications for the broader operational fleet. [1]

NEUROSYMBOLIC VLA REASONING

  • The STeP framework uses Signal Temporal Logic within a hierarchical VLA structure to give manipulation policies quantitative temporal and spatial guarantees on instruction following. [5]

MIT BERTSEKAS OBITUARY

  • Professor Emeritus Dimitri Bertsekas, whose foundational work on dynamic programming, optimal control, and large-scale computation shaped modern reinforcement learning and AI, died at age 83.

📐 STANDARDS & POLICY

NIST DIRECTOR CONFIRMED

  • Arvind Raman, former dean of engineering at Purdue University, was confirmed as the 18th Director of NIST as of July 6, 2026.

NIST MEP MANUFACTURING FUNDING OPPORTUNITY

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

NIST AI CYBERSECURITY GUIDELINES

  • Draft NIST guidelines published in December 2025 rethink cybersecurity for the AI era, helping organizations mitigate risks when incorporating AI into operations.

NIST CAISI AI AGENT SECURITY RFI

  • NIST's Center for AI Standards and Innovation published a Request for Information on securing AI agent systems, seeking input from industry and academia.

NIST AI IN MANUFACTURING CENTERS

  • NIST launched Centers for AI in Manufacturing and Critical Infrastructure in collaboration with MITRE Corporation as part of U.S. AI leadership efforts.

IEEE CERTIFAIED AI ETHICS CERTIFICATION

  • IEEE SA's CertifAIEd program is actively promoted as a professional development credential for responsible AI governance and ethics certification.

HIDDEN AI SAFETY FAILURES

  • A new arXiv perspective paper argues current AI safety discourse over-indexes on dramatic visible failures while neglecting quieter, deployed-system failures that are plausible rather than spectacular.

AGENTIC CI/CD PIPELINE ATTACK SURFACE

  • A paper studying a five-agent CI/CD pipeline across five production LLMs found agents verify injected malicious instructions but fail to act against them, exposing deployed pipelines to laundered-code attacks.

💰 FUNDING & PROGRAMS

NSF GENESIS MISSION ALIGNMENT

  • NSF Chief of Staff Brian Stone announced NSF is joining the White House OSTP, DOE, and federal partners to advance the Genesis Mission, a DOE-led national AI-driven scientific discovery platform anchored at ORNL and 17 national labs.

NSF $83M DATA INFRASTRUCTURE INVESTMENT

  • NSF announced $83 million in awards through the Integrated Data Systems and Services (IDSS) program to expand data infrastructure researchers can use alongside AI computing. [2]

NSF UNLOCKING DATASET VALUE PROGRAM

  • NSF launched the Unlocking Dataset Value for AI-Enabled Scientific Discovery program, a new investment targeting scientific community datasets to enable AI-driven discovery and innovation.

UKRI FIVE-YEAR STRATEGY

  • UKRI published an ambitious five-year roadmap to back more than 20,000 researchers, targeting breakthrough discoveries in AI, quantum, and other future technologies across the UK.

UKRI TWO NEW AI RESEARCH LABS

  • UKRI, via EPSRC, launched two new AI research labs to develop next-generation AI systems and secure the UK's position as a global AI leader, announced June 23, 2026.

NSF BRAIN-COMPUTER INTERFACE EXOSKELETON

  • An NSF podcast highlights researcher Payam Heydari's brain-computer interface that controls a robotic exoskeleton, targeting transformation of life for people with spinal cord injuries.

DARPA MXO SPARK TANK

  • DARPA's Multi X Office is running a Spark Tank and Pitch Day event to engage out-of-the-box innovators across multiple technical domains.

📄 RESEARCH

ADHESIVE FORCE PREDICTION FOR SOFT GRIPPERS

  • A deep learning model predicts time-resolved adhesive force trajectories for viscoelastic Hertzian contacts, replacing costly full numerical simulations with fast inference; directly applicable to soft robotic gripping and manipulation.

OPEN-VOCABULARY OBJECT NAVIGATION WITHOUT TRANSFORMERS

  • A linear attention policy outperforms Transformer-based backbones for open-vocabulary object goal navigation, delivering competitive performance under partial observability with substantially lower compute cost.

BEHAVIOR TREE SYNTHESIS FROM TEMPORAL LOGIC

  • A new method synthesizes correct-by-construction Behavior Trees directly from Signal Temporal Logic specifications, adding quantitative timing guarantees that Linear Temporal Logic synthesis cannot express; demonstrated on robotic missions.

VISUAL REINFORCEMENT LEARNING FOR QUADRUPED NAVIGATION

  • Predictive Training with Latent Imagination trains quadruped navigation policies to anticipate moving obstacle trajectories rather than react, reducing the lag that causes failures in dynamic environments.

QUANTIZATION ROBUSTNESS FOR EDGE ROBOTICS

  • Recti-Q shows that post-training quantization of large vision backbones preserves clean in-distribution accuracy but can drastically degrade out-of-distribution robustness in edge robotics perception pipelines, and proposes feature-space rectification to fix it.

MOEQUALITY IN MoE TRAINING MEMORY

  • SkewAdam (part of the SkewAdam study on tiered optimizer state allocation) demonstrates that for a 6.78B-parameter MoE model, AdamW's optimizer state alone consumes 50.6 GB versus 12.6 GB of weights, motivating tiered memory placement to make large MoE training practical.

📎 Sources

  1. DARPA, U.S. Air Force fly AI-controlled F-16 — DARPA News
  2. NSF announces $83M investment in integrated data systems and s… — NSF News
  3. DASH Robot: Minimalistic Design and Optimal Aerial-Terrestrial… — arXiv cs.RO (Robotics)
  4. New chip could help tiny robots traverse complex environments — MIT News — AI
  5. STeP: Signal Temporal Logic for Precise Specifications for Act… — arXiv cs.RO (Robotics)
  6. RoboInter1.5: A Holistic Intermediate Representation Suite for… — arXiv cs.RO (Robotics)
  7. Closing the Loop in Humanoid VLA: Persistent 3D Object Tokens … — arXiv cs.RO (Robotics)
  8. FM-VLA: Force-based Memory for Vision-Language-Action Models i… — arXiv cs.RO (Robotics)
  9. Eversion-based robots can enable safe access,steering and endo… — arXiv cs.RO (Robotics)
  10. MVP-Tac: A Miniaturized Dual-Modal Vision and Photoelastic Tac… — arXiv cs.RO (Robotics)

Curated from official sources — DARPA/NSF/NIST/IEEE/ORNL/MIT/UKRI/arXiv. Informational only.
Serial 20260723-00-v38 · 2026-07-23 00:01 UTC · pulse.uzylab.com

Read next