Sign in Subscribe
Robotics

🤖 Robotics Pulse · 2026-09-02 00:01 UTC

ROBOTICS PULSE

Wednesday, September 2, 2026

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

⚡ TL;DR

DARPA's AI-controlled F-16 under the VENOM program marks a historic milestone for autonomous combat aircraft, while today's feed is dominated by a flood of VLA model research pushing robot manipulation toward real-world generalization. [1]

Across 262 items, the mood is high-velocity and applied: embodied AI is moving fast from lab to deployment, and funding bodies are placing big structural bets.

🤖 ROBOTICS

HUMANOID AND LEGGED LOCOMOTION

  • Researchers trained a humanoid to traverse monkey bars by jumping, crossing sparse bar structures, and landing safely using agile whole-body motion learned via reinforcement learning. [2]
  • SleepWalking introduces privileged representation shaping so legged robots can navigate blind, without explicit estimation of hidden physical variables, outperforming observation-history baselines. [3]
  • Stay Seated teaches a humanoid to perform omnidirectional locomotion on a passive caster-wheel chair, offloading weight support during desk-work scenarios without active actuation. [4]
  • A dual-cam parallel elastic actuator with a shared gas-spring compensates torques in both pitch and roll for humanoid ankles, improving torque capacity and energy efficiency. [5]

MANIPULATION AND GRASPING

  • Motus2 is a self-evolving world model for dexterous manipulation that couples scene prediction and action in a closed loop, letting the robot improve from its own experience without offline resets. [6]
  • Zeva uses in-context causal learning so robots can adapt to unseen physical conditions on the fly during real deployment, updating manipulation knowledge from live interactions. [7]
  • CometVLA co-trains VLA models on an embodied data pyramid that layers physical commonsense QA, egocentric video, and robot trajectories, reducing brittleness in manipulation requiring physics intuition. [8]
  • Temporal Forcing adds 4D representation alignment to VLA models, injecting temporal cues to resolve observation aliasing between visually similar states in long-horizon manipulation. [9]
  • SUN (Persistent Programs) bridges model-based control and learned policies by retaining task semantics through a language-grounded programming layer, eliminating hand-crafted rewards. [10]
  • A training-free suction grasping system for deformed aseptic cartons decouples target identification via open-vocabulary VLM from grasp-point scoring via geometric surface analysis.
  • Picking Bins Empty presents a hierarchical hybrid bin-picking method with online self-learning of grasp points, breaking deadlocks that trip up purely model-based systems at full bin clearance.
  • ChainSplat uses a physics-inspired screw-theoretic model to learn deformable linear object dynamics (cables, ropes, hoses) from multi-view RGB video for robotic manipulation.
  • Proximity3D reconstructs 3D shape from a curved capacitive textile sensor treated as a non-planar sensing manifold, enabling touch-based geometry estimation without cameras.
  • SpectraTac is a compact, camera-free optical tactile sensor using distributed color sensing that delivers rich tactile data at low cost and computational overhead.
  • Anomaly detection on sub-centimetre industrial components is demonstrated using vision-based tactile sensors that directly capture fine surface geometry where optical cameras fall short.
  • Contact-Guided Exploration for non-prehensile locomanipulation uses multi-critic RL to handle complex hybrid contact dynamics and sparse reward signals for moving heavy objects.

NAVIGATION AND AUTONOMOUS VEHICLES

  • SmoothRL enables online RL during asynchronous execution on physical robots, reconciling the tension between real-time smooth motion and the precision demands of generalist policies.
  • LightNav-0 elicits spatial intelligence from pre-trained VLMs for generalist embodied navigation across heterogeneous goals, environments, and robot embodiments with no task-specific training.
  • Self-Aware Active Learning equips autonomous driving systems with uncertainty estimation to identify distribution shifts and long-tail events, triggering targeted data collection for continual improvement.
  • CanonNav disentangles navigation behavior from camera geometry in visual navigation, enabling cross-platform knowledge transfer without re-training for each sensor configuration.
  • Driving on Memory introduces a memory-augmented end-to-end driving model evaluated on NAVSIM and Bench2Drive with simulation-based safety metrics rather than trajectory imitation.
  • CAVE-NAV is a VLM-based autonomous 3D navigation system designed for underwater cave environments where dense visual features are unavailable due to visual degradation.
  • A year-long study of autonomous navigation in subarctic forests documents challenges including GNSS denial under tree canopies and unreliable cloud connectivity, providing empirical failure analysis.
  • HorizonNet uses a CNN to extract horizon lines from 360-degree panoramic images for position estimation of unmanned surface vessels in coastal and archipelago regions.

AERIAL AND SPACE ROBOTICS

  • DARPA and the U.S. Air Force flew an AI-controlled F-16 under the VENOM program, demonstrating scalable AI development for operational fleet deployment in a historic milestone. [1]
  • DARPA's Mission Robotic Vehicle has launched and is en route to geosynchronous orbit under the Robotic Servicing of Geosynchronous Satellites program, targeting in-orbit servicing.
  • A tilt-rotor bicopter drone design is optimized for both static and dynamic stability using a modular architecture that combines fixed-wing endurance with rotary-wing VTOL capability.
  • Research characterizes the merging turbulent jet structure beneath a hovering quadrotor using high-resolution measurement, with implications for multi-rotor formation spacing.

MEDICAL AND WEARABLE ROBOTICS

  • MiBOT is a head-worn soft robot that delivers human-like massage motions to modulate cardiovascular responses, addressing a gap where prior robotic massage systems focused only on torso and limbs.
  • Contrast-free autonomous navigation of untethered endovascular microrobots uses single-plane fluoroscopy and 3D pose inference, removing dependence on contrast agents for vascular procedures.
  • Data-Centric Neuromotor Interfaces use flexible material sensors to decode movement intention on edge devices, enabling dexterous portable human-machine interaction.

MULTI-ROBOT SYSTEMS

  • Provably safe decentralized contingency MPC for nonlinear multi-agent systems guarantees recursive feasibility and Lyapunov-type convergence under limited sensing and plug-and-play operation.
  • Coordinated motion planning for multi-arm systems uses iterative LQ games to balance scalability and safety in shared workspaces for high-DOF manipulators.
  • Cooperative Risk-Aware Exploration uses algorithmic altruism in a game-theoretic framework to distribute hazard across heterogeneous multi-robot teams during dangerous environment mapping.
  • Probabilistic multi-robot gas source localization operates with uncalibrated heterogeneous sensors via a distributed estimation approach, avoiding centralized fusion that requires sensor calibration.
  • CIG-RL combines curiosity-driven and information-guided RL for source term estimation of hazardous gas releases using mobile sensor platforms in uncertain environments.

SIMULATION AND TRAINING DATA

  • MIT's SceneSmith uses collaborative AI agents to generate realistic 3D environments such as kitchens and hotel rooms, producing robot training data without physical collection.
  • EMERGE-Policy organizes a robot's capabilities as a graph-structured agentic framework coordinating perception, reasoning, prediction, action, verification, and memory rather than a single policy.
  • DARP is a calibrated dual-arm RGB-D-IR dataset for object-centered robotic perception from two independent viewpoints, addressing occlusion and incomplete surface visibility.
  • Behavior-Skill is a fine-grained benchmark for evaluating VLA policies on long-horizon mobile manipulation tasks at the level of constituent skills, not just final success.
  • CoCoBench introduces a cooperative coordination benchmark for embodied multi-agent task planning with fine-grained diagnostics beyond aggregate task completion rates.
  • VR-based remote human-robot interaction is evaluated for greenhouse leaf inspection and soil moisture assessment using an unmanned ground vehicle with a manipulator arm.

INDUSTRIAL AND FIELD ROBOTICS

  • MaCoPlanner is an LLM-assisted task planning framework for robotic industrial panel operation that compiles safety rules, operating procedures, and device-state constraints from heterogeneous manuals.
  • PanelShield adds verifiable closed-loop safety planning on top of foundation-model planners for industrial panel operation, providing computable constraint guarantees missing from semantic-only systems.
  • GAFT applies geo-anchored fine-tuning to identify terrain hazards like high-centering from rare failure events in off-road navigation, addressing the data scarcity of costly field failures.
  • Failure or Drift evaluates monocular SLAM under synthetic and real-world corruptions including adverse weather, blur, and illumination changes, testing whether synthetic stress tests predict real failures.

🧠 AI & MODELS

LARGE LANGUAGE AND VISION-LANGUAGE MODELS

  • Soft Latent Thinking replaces the LM head during reasoning with a learned soft representation, allowing chains of thought in continuous latent space and cutting discrete token overhead.
  • LOCI (Locator-Critic with Refinement Loop) argues VLM failures stem from poor localization of critical image details, not high-level reasoning, and adds a locator-critic loop to fix grounding.
  • LLM self-modeling is benchmarked on verifiable behavioral questions such as whether a prompt edit changes the final answer, revealing systematic gaps between stated and actual model behavior.
  • Knowledge-aligned SFT constrains fine-tuning targets to facts already robustly internalized by the base model, framing hallucination as a mismatch between training targets and model knowledge.
  • Sycophantic agreement is shown to transfer via contrastive preference optimization even on neutral data, revealing a training mechanism by which models learn to over-affirm users.
  • LLM judges are shown to verify presence of information well but exhibit omission blindness, consistently failing to flag information that was established in a clinical encounter but missing from the note.
  • An audit of three commercial AI scribes across 565 notes from 142 UK primary-care and US ambulatory consultations finds roughly one note in three contains a significant omission.
  • A four-stage black-box protocol for identifying anonymously released frontier AI models is introduced, motivated by the wave of stealth codename launches observed in 2025 and 2026.
  • Scaling large reasoning models beyond human supervision examines how RLVR can extend from math and code to open-ended agentic tasks where automatic reward verification is unavailable.
  • On-policy distillation is re-examined: because the teacher scores student-generated off-policy trajectories, teacher reliability degrades and self-improvement may outperform noisy teacher signals.
  • VISTA introduces verifier-informed student-to-teacher adaptation for on-policy self-distillation, letting the teacher distribution evolve alongside the student rather than remaining fixed.
  • Reconciling Process Supervision with Outcome-Based Credit addresses coarse trajectory-level advantage assignment in long-horizon agentic RL by layering on-policy self-distillation for finer credit.

EFFICIENCY AND TRAINING METHODS

  • Normalized Low-Rank Adaptation (LoRA) addresses unstable early optimization from zero-initialized up-projection by adding normalization to regularize training dynamics.
  • Sliding-window attention is shown to outperform linear attention alternatives for LLMs in memory and throughput benchmarks, challenging the assumption that linear attention is the successor to quadratic.
  • A Universal Context-Reuse Layer enables cross-model KV cache sharing so that multiple LLMs serving repeated contexts avoid redundant prefill computation.
  • CE-MoE reduces all-to-all communication overhead in Mixture-of-Experts training through heterogeneous layer reconfiguration, cutting the collective cost that dominates distributed MoE training time.
  • Deriving scaling laws for the OpenEuroLLM models characterizes optimal learning rate and batch size as functions of model capacity and data scale during dense LLM pretraining.

REASONING AND AGENTS

  • CAER (Causal Action Effect Reweighting) trains world models by upweighting action-relevant regions of the loss rather than applying uniform mean squared error across space and time.
  • Autonomous scientific research agents are improved by automatic rubric induction that generates task-specific evaluation criteria before the agent attempts open-ended research workflows.
  • Token-Efficient Data Reasoning Agents reduce prohibitive token costs of LLM agents reasoning over unstructured enterprise documents via adaptive structuring before inference.
  • MNIST-PRO recasts MNIST as a partially observable world to isolate and evaluate the perceptual-state construction capability of AI agents that must coordinate active sensing with working memory.
  • Learning action models with conditional and quantified effects is addressed by Online Hypothesis-Driven Conditional Action Model Learning, which uses uncertainty-guided exploration to remain tractable.

PHYSICS-AWARE AND SCIENTIFIC AI

  • GeoPT teaches AI models to apply basic physics principles, enabling more efficient and accurate simulation of how objects respond to wind and water across a wider range of scenarios.
  • Rotational equivariance in machine learning on 3D data is covered in a comprehensive tutorial addressing coordinate-frame independence for physics, materials science, and 3D vision applications.
  • Singular curvature in ReLU training proves that at fixed non-resonant step sizes, automatic differentiation exactly differentiates the hard-ReLU GD program, but gradient flow and GD need not commute after differentiation.
  • Real-time neural network emulators for plasma shape control in tokamaks are experimentally demonstrated on MAST Upgrade, replacing offline linearized virtual circuits with online neural surrogates.

📐 STANDARDS & POLICY

  • IEEE Standards Association publishes a framework defining what makes a system both autonomous and intelligent, covering AIS applications across healthcare and transportation.
  • IEEE's age verification framework incorporates the 5Rights principles for ethical digital design for children, with an associated online certification program for platforms.
  • NIST has joined the National Genesis Mission, executing two efforts through its Centers for AI in Manufacturing and Critical Infrastructure to accelerate applied AI innovation.
  • New NIST Director Arvind Raman, formerly dean of engineering at Purdue University, was confirmed as the 18th director in July 2026.
  • Stress-testing responsible AI benchmarking finds that cheaper evaluation protocols can shift benchmark conclusions for dense and mixture-of-experts models, undermining reliability of cost-saving audits.
  • VFR-Audit introduces verdict-level reliability analysis for fairness audits in clinical AI, addressing the instability of pass-or-fail governance verdicts repeated across hospital sites over time.

💰 FUNDING & PROGRAMS

  • NSF announces over $1.5 billion across 12 new notices of funding opportunity for foundational research including basic and use-inspired inquiry, targeting American technological leadership.
  • NSF invests $108 million in six advanced materials science research centers to push beyond the state of the art in scientific frontiers including exotic material properties.
  • NSF launches three new Science and Technology Centers with a $90 million five-year investment in transformative research to advance American leadership in science and technology.
  • NSF State and Regional AI Infrastructure Hubs initiative expands access to compute for researchers, students, and educators through regional partnerships among governments, academia, and industry.
  • NSF invests $47 million over five years in a pilot initiative for four-year PhD programs with real-world industry research placements at universities across the U.S.
  • UKRI launches two new AI research labs backed by EPSRC to develop next-generation AI systems and secure the UK's position as a global leader in artificial intelligence.
  • UKRI publishes its five-year roadmap targeting breakthroughs in AI and quantum, pledging support for more than 20,000 researchers to drive growth across the UK.
  • King Charles III officially opened the UK Space and Defence Gateway at Harwell Science and Innovation Campus including RAL Space, operated by STFC, on 10 July 2026.
  • Innovate UK announces its largest-ever Women in Innovation cohort, backing 100 women founders across manufacturing, digital tech, and life sciences.

📄 RESEARCH

PAPER 1 - LEARNING FROM WHOLE-ARM SOFT ROBOT INTERACTION

Inspired by elephants and octopuses, researchers trained a soft robot to simultaneously infer object properties and manipulate objects through distributed whole-arm contact, treating sensing and acting as a single inseparable process. The system learns representations from large-area compliant interactions rather than fingertip contact alone.

PAPER 2 - AUTONOMOUSLY ACQUIRING MANIPULATION SKILLS WITH LANGUAGE-DRIVEN QUALITY-DIVERSITY

A new method applies quality-diversity algorithms to robot manipulation, using language models to automatically write the success conditions and diversity metrics that normally require expert designers, enabling robots to autonomously build diverse motion primitive libraries without human-specified rewards.

PAPER 3 - SCAFFOLDING FOUNDATION MODELS FOR LONG-HORIZON NAVIGATION

The paper argues that VLMs and low-level controllers have complementary, non-overlapping strengths for physical-world navigation. A scaffolding framework combines VLM high-level inference with reliable closed-loop behavior to push long-horizon navigation beyond what either component achieves alone.

PAPER 4 - LEARNING TO INFER SOFT ROBOT MANIPULATION THROUGH WHOLE-ARM CONTACT

SmoothRL addresses a core deployment gap: state-of-the-art robot policies are typically trained assuming synchronous execution, but real hardware runs asynchronously. The method performs online RL during asynchronous execution without pausing for policy updates, maintaining smooth real-time motion.

PAPER 5 - REPRESENTATION-CENTRIC CONTINUED PRE-TRAINING FOR VLA MODELS

Rather than scaling robot trajectory data, this work targets representation quality as the central bottleneck in VLA models, proposing continued pre-training methods that improve how the backbone encodes physical-world structure without requiring more costly embodied data collection.

End of edition. Sources: DARPA, NSF, NIST, IEEE SA, UKRI/EPSRC, MIT News, arXiv cs.RO, cs.AI, cs.LG.

📎 Sources

  1. DARPA, U.S. Air Force fly AI-controlled F-16 — DARPA News
  2. Learning Agile Perceptive Traversal of Sparse 3D Structures fo… — arXiv cs.RO (Robotics)
  3. SleepWalking: Privileged Representation Shaping for End-to-End… — arXiv cs.RO (Robotics)
  4. Stay Seated: Learning Omnidirectional Humanoid Locomotion on a… — arXiv cs.RO (Robotics)
  5. A Dual-Cam Parallel Elastic Actuator with Shared Gas-Spring Co… — arXiv cs.RO (Robotics)
  6. Motus2: A Self-Evolving General World Model for Dexterous Mani… — arXiv cs.RO (Robotics)
  7. Zeva: In-Context Causal Learning for Generalizable Embodied Ma… — arXiv cs.RO (Robotics)
  8. CometVLA: Co-Training on an Embodied Data Pyramid towards Phys… — arXiv cs.RO (Robotics)
  9. Temporal Forcing: 4D Representation Alignment for Vision-Langu… — arXiv cs.RO (Robotics)
  10. SUN: Persistent Programs For Language-Grounded Control-to-Lear… — arXiv cs.RO (Robotics)

Curated from official sources — DARPA/NSF/NIST/IEEE/ORNL/MIT/UKRI/arXiv. Informational only.
Serial 20260902-00-v66 · 2026-09-02 00:01 UTC · pulse.uzylab.com

Read next