Sign in Subscribe
Robotics

🤖 Robotics Pulse · 2026-07-15 00:01 UTC

ROBOTICS PULSE

July 15, 2026

⚡ TL;DR

MIT's JARVIS Challenge put AI copilots to the ultimate engineering test - designing, building, and testing a jet engine - raising pointed questions about AI's real ceiling in tough-tech development. [1] Today's briefing is dense with manipulation and locomotion breakthroughs from arXiv, a major NSF regional innovation funding round, and a flurry of embodied AI advances pushing toward open-world physical intelligence.

🤖 ROBOTICS

FLOATFORM SWARM ROBOTICS

  • MIT researchers unveiled FloatForm, a swarm of small aquatic robots that snap together like ants forming a raft, self-assembling into reconfigurable floating structures on water surfaces. [2]

WHOLE-BODY LOCO-MANIPULATION

  • PAKE (Partial Kinematic Embeddings) addresses high-precision whole-body loco-manipulation by decomposing the high-dimensional action space of many-DoF robots using partial kinematic embeddings, reducing the complexity of exploiting full-body redundancy. [3]

VLA WITH ROBOT-CENTRIC 3D VISION

  • "See like a Robot" paper introduces robot-centric pointmaps for vision-language-action (VLA) models, correcting the frame mismatch between camera-space observation and robot-frame action prediction to improve manipulation accuracy. [4]

DEXTEROUS TELEOPERATION

  • A new paper on human-level dexterous teleoperation enables robots to replicate in-hand manipulation including tool use, grasp swaps, and finger gaiting - the contact transitions humans perform routinely but robots struggle to match. [5]

HUMANOID DEXTEROUS MANIPULATION VIA RETARGETING

  • Researchers tested whether the retarget-then-RL recipe from whole-body locomotion transfers to dexterous manipulation, finding the combination requires careful adaptation given the contact-rich, precision-sensitive nature of hand tasks. [6]

BIMANUAL MOBILE MANIPULATION

  • Mixture of Frames Policy denoises actions in multiple coordinate frames simultaneously for bimanual mobile robots, handling tasks where local end-effector frames and global base frames are both needed within one diffusion policy. [7]

BIPEDAL LOCOMOTION ON FLOWABLE TERRAIN

  • New work demonstrates robust bipedal robot locomotion on sand and other flowable slopes by modeling foot-driven terrain manipulation, explicitly accounting for how the foot reshapes the surface underfoot during each step. [8]

LEGGED ROBOT LOCOMOTION WITH KOOPMAN OPERATORS

  • SKooP encodes left-right locomotion symmetry via Symmetric Koopman Predictions, improving sample efficiency and generalization for legged RL policies beyond standard low-dimensional benchmarks. [9]

CONTINUUM ROBOT KINEMATICS ON GPU

  • CR-Solver is a GPU-accelerated kinematics solver for tendon-driven continuum robots, dramatically speeding up motion planning in confined and unstructured environments where these robots are deployed. [10]

DARPA RSGS SATELLITE SERVICING ON TRACK FOR LAUNCH

  • DARPA's Robotic Servicing of Geosynchronous Satellites (RSGS) program is approaching its most significant milestone with a planned launch in 2026, bringing in-orbit robotic satellite repair closer to operational reality.

ENDOVASCULAR INTERVENTION ROBOT

  • A compact top-loading robot for endovascular interventions was evaluated for procedural coverage, ergonomics, and precision, targeting stroke and vascular disease treatments that currently require specialist surgeons.

QUADRUPED INSPECTION WITHOUT MAPPING

  • "Stop to Decide" introduces latency-aware proprioceptive navigation primitives for a Jetson Orin-constrained quadruped running its stair-detection loop at roughly 15 Hz, enabling reliable inspection without a pre-built map.

INDOOR UAV SWARM NAVIGATION

  • A mission-oriented UAV swarm framework uses sketch-based spatial priors to coordinate indoor inspection and patrol flights, converting region-level mission descriptions directly into coordinated multi-UAV trajectories.

SCENESMITH FOR ROBOT TRAINING DATA

  • MIT's SceneSmith system uses collaborative AI agents to auto-generate realistic 3D environments including kitchens, hotels, and living rooms, giving robots scalable simulated training data for everyday chores.

WALA: LEARNING FROM ACTION-FREE VIDEO

  • WALA learns executable latent robot actions from a combination of action-labeled demonstrations and unlabeled human or robot videos, addressing the bottleneck of expensive action-annotation at scale.

LUMO-2 WORLD-ACTION MODEL

  • Lumo-2 is a latent world-action model that reasons over world dynamics in latent space before generating actions, aiming for predictive, aligned, and scalable robot learning across novel tasks.

XIAOMI ROBOTICS WORLD FOUNDATION MODEL

  • Xiaomi-Robotics-U0 adapts large image and video generation foundation models to embodied scenarios by enforcing multi-view consistency, geometric coherence, and robot kinematic constraints during synthesis.

SOCIAL ROBOT TRIGGERED BY NONVERBAL CUES

  • A retail store deployment study found a social robot system using nonverbal cue detection (approach, wave, point, show item) increased proactive customer engagements well past the 15 percent baseline for speech-triggered interactions.

TACTILE BENCHMARK FOR DEXTEROUS MANIPULATION

  • TactiDex introduces a real-world tactile-guided benchmark for human-like dexterous manipulation, going beyond kinematic retargeting to include contact formation, force regulation, and stable grasp metrics.

WHOLE-ARM TACTILE AND VISION CONTROL

  • A new contact-centric control framework uses both tactile sensing and vision for whole-arm manipulation, handling the sliding and breaking contacts that defeat standard end-effector-centric pipelines.

AI COPILOTS FOR JET ENGINE DESIGN

  • MIT's JARVIS Challenge had students design, build, and test a jet engine with AI copilots, generating concrete data on where AI assistance accelerates and where it falls short in high-performance aerospace engineering. [1]

CORAL-AUV REINFORCEMENT LEARNING CONTROL

  • CORAL-AUV uses CFD-informed reinforcement learning for fine-grained positioning and control of autonomous underwater vehicles, improving robustness over traditional methods when vehicle configuration or environment changes.

🧠 AI & MODELS

MACH-MIND-4-FLASH 35B MOE MODEL

  • Mach-Mind-4-Flash is a 35B-parameter Mixture-of-Experts agentic model activating only 3B parameters per token; through post-training optimization alone it claims performance matching or exceeding 100B-parameter-class models.

SOOFI S 30B HYBRID MAMBA-TRANSFORMER

  • Soofi S 30B-A3B is an open-source sovereign MoE hybrid Mamba Transformer for German and English, activating 3B of 30B parameters per token with a near-constant inference cache as context grows.

MURAKKAB AI AGENT EFFICIENCY SYSTEM

  • MIT's Murakkab optimizes the design and deployment of multistep AI agent workflows, targeting speed and energy consumption in applications that chain multiple LLM calls and tool uses.

DEVAVRAT SHAH ON REAL-WORLD AI DECISION-MAKING

  • MIT Professor Devavrat Shah is developing methods for constant decision-making under limited compute, bridging the gap between powerful AI models and the constrained real-world systems that must run them.

INTERACTION SCALING AS A THIRD AXIS OF TEST-TIME COMPUTE

  • A new paper proposes "interaction scaling" - using external feedback loops rather than longer internal reasoning chains or more samples - as a fundamentally distinct third axis for spending test-time compute.

HOURGLASS REASONING FOR INDUCTIVE LLM TASKS

  • "Think Through a Bottleneck" shows that structurally enforced stage isolation in prompting, not just explicit rule verbalization, is what actually strengthens few-shot inductive reasoning in LLMs.

DISTRIBUTED BACKDOORS IN MULTI-AGENT LLM SYSTEMS

  • Researchers demonstrate a critical security gap: a harmful payload split across agents passes every local safety monitor individually, yet assembles into a complete attack, defeating standard per-message safety checks.

VINCI2 PROACTIVE EGOCENTRIC VIDEO ASSISTANT

  • Vinci2 uses continuous egocentric video to provide proactive assistance without waiting for user queries, deciding autonomously when to speak up based on evolving visual context.

ASSOCIATIVE RECURRENT MEMORY TRANSFORMER FOR LONG CONTEXT

  • The Associative Recurrent Memory Transformer (ARMT) addresses LLM context length limits by replacing quadratic attention with associative recurrent memory, keeping compute and memory scaling practical.

AGENT HACKS AGENT: AUTOMATED RED-TEAMING

  • A new framework called "Agent Hacks Agent" performs automated red-teaming of production LLM agents including Claude Code and Codex, targeting safety failures over untrusted files, commands, and workspace state.

CLAP: DIRECT VLM-TO-VLA ADAPTATION

  • CLAP converts pretrained vision-language models directly to vision-language-action policies via language-action grounding, without the large-scale robot post-training that typically erodes the original VLM's semantic backbone.

METACOGNITION IN LLMS SURVEY

  • A comprehensive survey maps the state of metacognition in LLMs - covering self-monitoring, uncertainty awareness, and self-correction - framing it as a cornerstone of capable and transparent AI systems.

NIST AI EVACUATION MODEL

  • A NIST-led team created an AI model that identifies safe fire evacuation routes step-by-step for single-story floor plans, with a multilevel version under development.

NIST MATHEMATICAL PROOF FOR CONTINUOUS AI SECURITY MONITORING

  • NIST published a mathematical proof, extending Goedel's incompleteness logic, supporting a shift from static AI security audits to continuous monitor-and-update models for deployed AI systems.

📐 STANDARDS & POLICY

NIST AI AGENT STANDARDS INITIATIVE

  • NIST launched the AI Agent Standards Initiative to ensure next-generation AI agents can be widely adopted with confidence, operate securely on behalf of users, and interoperate smoothly across the digital ecosystem.

NIST AI CONSORTIUM EXPANSION

  • NIST expanded its AI consortium's scope and called for new members, organizing work into six task groups focused on different aspects of AI measurement science and evaluation.

IEEE AT GENEVA DIGITAL WEEK

  • IEEE participated in Geneva Digital Week (6-10 July 2026), engaging in global digital governance discussions with governments, international organizations, and industry on the future of digital technologies.

NIST SBIR AI FUNDING

  • NIST allocated over 3 million dollars to eight small businesses in seven states under the SBIR program, with AI among the targeted technology areas alongside biotechnology, semiconductors, and quantum.

💰 FUNDING & PROGRAMS

NSF REGIONAL INNOVATION ENGINES

  • NSF awarded 12 new Regional Innovation Engines to teams spanning 20 U.S. states, building and scaling regional innovation clusters intended to accelerate research, create jobs, and drive economic growth.

UKRI 2025-2026 ANNUAL REPORT

  • UKRI published its 2025-2026 annual report, highlighting advances ranging from new cancer treatments to plant-based plastic replacements as evidence of UK research driving knowledge, health, and economic growth.

UK SPACE AND DEFENCE GATEWAY OPENS

  • King Charles III officially opened the UK Space and Defence Gateway at the Harwell Science and Innovation Campus on 10 July 2026, including RAL Space operated by STFC.

DARPA AI FORGE

  • DARPA's AI Forge initiative released a new report and RFI aimed at aligning government, academia, and industry around forward-looking AI research priorities for national security applications.

DARPA LIFT CHALLENGE

  • DARPA invited the first wave of competitors to the Lift Challenge, with 6.5 million dollars in prizes available across the first tranche of selected teams.

DARPA YOUNG FACULTY AWARDS 20TH ANNIVERSARY

  • DARPA celebrated 20 years of its Young Faculty Award program, which has supported over 500 rising research stars from more than 60 institutions, and announced new Director's Fellows.

📄 RESEARCH

ENVIRONMENT-DEPENDENT ACTION REPRESENTATIONS (EDAR)

  • EDAR learns action representations conditioned on the robot's specific environment rather than treating actions as environment-agnostic, reducing noise and redundancy in control trajectories for manipulation tasks.

PIX2ACT: IMAGE-SPACE MANIPULATION POLICIES

  • Pix2Act represents manipulation actions as 2D trajectories in camera image space and uses equivariant augmentation to handle out-of-frame trajectories and precision limitations, bridging compact representation with complex 3D policy learning.

WARPMPC: GPU-ACCELERATED MODEL PREDICTIVE CONTROL

  • WarpMPC solves batches of 10,000 to over 100,000 sequential quadratic programming iterations simultaneously on GPU using ADMM with unrolled LDL factorization, enabling high-throughput MPC for real-time robot control.

GEOGSSLAM: MONOCULAR 3D RECONSTRUCTION

  • GeoGS-SLAM combines 3D Gaussian Splatting with geometric priors from pretrained feed-forward models to achieve competitive tracking and mapping from a single monocular camera, eliminating the need for external depth sensors.

NEURAL ACTUATOR MODELING FOR SIM-TO-REAL

  • NeuralActuator learns a continuous-time neural surrogate of actuator dynamics on low-cost robot platforms where the standard linear current-to-torque model breaks down, directly reducing one of the largest remaining sources of sim-to-real error in learned robot policies.

📎 Sources

  1. Can AI build a jet engine? JARVIS Challenge tests role of AI c… — MIT News — AI
  2. Tiny robot boats build floating structures — MIT News — AI
  3. PAKE: Learning Whole-Body Loco-Manipulation with Partial Kinem… — arXiv cs.RO (Robotics)
  4. See like a Robot: Robot-Centric Pointmaps for Vision-Language-… — arXiv cs.RO (Robotics)
  5. Towards Human-level Dexterous Teleoperation — arXiv cs.RO (Robotics)
  6. A Minimalist Retargeting-Guided Reinforcement Learning Recipe … — arXiv cs.RO (Robotics)
  7. Mixture of Frames Policy: Multi-Frame Action Denoising for Bim… — arXiv cs.RO (Robotics)
  8. Robust bipedal locomotion on flowable slopes via foot-driven t… — arXiv cs.RO (Robotics)
  9. SKooP: Symmetric Koopman Predictions for Faster and More Gener… — arXiv cs.RO (Robotics)
  10. CR-Solver: GPU-Accelerated Kinematics Solver for Tendon-driven… — arXiv cs.RO (Robotics)

Curated from official sources — DARPA/NSF/NIST/IEEE/ORNL/MIT/UKRI/arXiv. Informational only.
Serial 20260715-00-v30 · 2026-07-15 00:01 UTC · pulse.uzylab.com

Read next