Results for "Reinforcement Learning"

50 articles found

Why robotics teams need virtual gyms before deployment

The article discusses how virtual simulation environments ('virtual gyms') are becoming essential for robotics development, combining digital twins, reinforcement learning, and sensor modeling to train robots safely before real-world deployment. It explains the sim-to-real gap as both a technical and production challenge, and outlines how selective fidelity in virtual environments can help robots learn to handle variability and edge cases without expensive or risky physical testing.

Robotics Business Review2d ago

How Path Robotics uses AI to optimize robotic welding

Path Robotics uses AI-driven vision guidance and torch path identification to optimize robotic welding operations in manufacturing environments, with the company also deploying Boston Dynamics Spot robots for mobile welding applications in shipbuilding. The article features an interview with Path Robotics CEO Andy Lonsberry discussing how physical AI addresses setup and operational challenges in robotic welding, alongside research perspectives from UC San Diego's Michael Yip on data-efficient robot control and reinforcement learning applications.

Robotics Business Review3d ago

Capturing token IDs during agentic interactions for better reinforcement learning

A new Rust-based proxy called Turnstile captures exact token-level information during agent-model interactions to improve reinforcement learning training. The tool sits between an agent harness and model backend to record precise token IDs and trajectories, solving the problem of data loss in text transcripts that prevents accurate RL parameter updates for language models performing complex multi-step tasks.

Amazon Science4d ago

Boston Dynamics brings its legged robots to the FIFA World Cup

Boston Dynamics' Atlas humanoid robot demonstrated advanced soccer skills at the FIFA World Cup, including the complex Ghost Rabona kick, as part of Hyundai's expanded partnership with FIFA. The company discussed how reinforcement learning and motion capture techniques enabled Atlas to learn complex object manipulation and athletic behaviors more efficiently than traditional model-based control methods.

Robotics Business Review7/6/2026

Humanoid says KinetIQ Ascend reinforcement learning approaches human-level dexterity

Humanoid, a London-based robotics startup, introduced KinetIQ Ascend, a reinforcement learning framework designed to enable humanoid robots to achieve 99.9% manipulation reliability at human speed and beyond. The company demonstrated the system across multiple industrial tasks including bin picking and bimanual handling, achieving throughput improvements of 42-85% and success rates exceeding 98% after only days of training, marking a shift from manual tuning to scalable, real-world deployment.

Robotics Business Review7/5/2026

Insights behind Kinisi's acquisition by Bear Robotics

Bear Robotics has acquired Kinisi Robotics, a Bristol-based company founded by Bren Pierce that uses reinforcement learning to automate industrial tasks. Kinisi's KR1 prototype leverages physical AI to bridge the gap between AI and the physical world, automating dirty and repetitive tasks alongside humans in industrial environments.

The Robot Report6/29/2026

Soft, robotic cells from morph embed physical AI into hardware

morph, a London-based robotics startup, is developing soft robotic cells that embed physical AI directly into deformable materials, enabling real-time sensing, adaptation, and morphological changes. Founded by Dr. Jean Nehme, the company integrates reinforcement learning with physics-based simulation to create adaptive materials inspired by octopuses that can learn continuously and respond to their environment, positioning soft robotics as a safer and more scalable alternative to rigid systems for human-robot interaction.

Robotics Business Review6/30/2026

We know how to build smarter robots. Now, we need to learn smarter ways to test them

The article discusses the gap between rapidly advancing autonomous robot capabilities and the lack of standardized testing and safety validation frameworks to evaluate them. The author proposes a five-level taxonomy for classifying robots by their control architecture—from teleoperation to full reinforcement learning autonomy—and argues that testing methodologies must evolve to handle increasingly autonomous systems, moving from simple test enumeration to formal safety guarantees and adversarial robustness evaluation.

The Robot Report6/27/2026

Autonomique deploys semi-humanoid robots and AI at Canadian Tier 1

Autonomique, an AI startup spun out of SRI International, is deploying its physical AI platform and mobile manipulator at Canadian automotive supplier F&P Manufacturing to address labor shortages and production complexity. The company's hardware-agnostic software combines teleoperation systems, generalist-specialist AI architecture, and reinforcement learning to enable industrial robots to adapt to multi-step workflows with human-like dexterity while maintaining manufacturing precision and reliability.

Robotics Business Review6/17/2026

Bike robot lands first unassisted front flip thanks to Ph.D. student

A Georgia Tech Ph.D. student developed a novel reinforcement learning method called iterative motion imitation (IMI) that enabled a bicycle robot (ultra-mobility vehicle) from the Robotics and AI Institute to perform the first unassisted acrobatic front flip. The technique teaches the robot to land gracefully by iteratively refining imperfect trajectories through simulation and imitation learning, addressing the challenge of absorbing impact forces without damage.

Phys.org Robotics6/10/2026

NVIDIA Enables the Next Era Of Physical AI Research With Agent Skills For Autonomous Vehicles, Robotics And Vision AI

NVIDIA unveiled new physical AI agent skills at CVPR designed to accelerate research in autonomous vehicles, robotics, and vision AI by automating workflows like scene reconstruction, synthetic scenario generation, and policy training. The announcement includes NVIDIA Cosmos 3, an open foundation model for physical AI, along with specialized tools like Neural Reconstruction skills, AlpaGym reinforcement learning framework, and Alpamayo 2 Super, a 32-billion-parameter vision-language-action model for autonomous driving development.

NVIDIA Blog AI6/3/2026

LLMs help robots understand vague instructions and focus on key details

MIT researchers have developed Masked Inverse Reinforcement Learning (Masked IRL), a system that uses large language models to help robots understand vague instructions and focus on relevant details from demonstrations. The approach reduces the need for extensive human demonstrations by 80% while enabling robots to safely complete manipulation tasks in warehouses, offices, and homes by automatically clarifying ambiguous prompts and identifying crucial environmental constraints.

Phys.org Robotics6/2/2026

Table tennis robot defeats some of world's best players - why this has major implications for robotics

Sony AI's table tennis robot Ace defeated some of the world's best players in recent evaluations, winning 3 out of 5 matches against elite competitive athletes, marking a significant advancement in AI systems operating in fast, uncertain real-world environments. The robot uses event-based vision sensors, high-speed cameras, and deep reinforcement learning to handle complex challenges like spin estimation and real-time ball tracking, demonstrating progress beyond controlled digital environments where AI has traditionally excelled.

Robohub5/18/2026

Sony AI table tennis robot outplays elite human players

Sony AI has developed Ace, a robot that defeats elite human table tennis players in competitive matches, marking the first time an autonomous robot has outplayed elite athletes in a physical sport. Ace combines event-based vision sensors, model-free reinforcement learning, and high-speed robotic hardware to achieve real-time perception, decision-making, and control at millisecond speeds. The research, published in Nature, demonstrates significant advances in physical AI agents capable of rapid adaptation and precise execution in complex, unpredictable real-world interactions.

Robohub4/22/2026

AGIBOT launches Genie Studio Agent zero-code application platform for robots

AGIBOT launched Genie Studio Agent, a zero-code application platform designed to simplify robot development and deployment at scale without requiring coding expertise. The platform features no-code workflow orchestration, simulation-first deployment, real-world reinforcement learning optimization, and comprehensive monitoring—addressing the critical challenge of efficiently deploying embodied AI robots in real-world environments.

Robotics Business Review4/13/2026

AI system learns to keep warehouse robot traffic running smoothly

MIT researchers and Symbotic developed an AI system using deep reinforcement learning that optimizes warehouse robot traffic by dynamically prioritizing which robots move first to avoid congestion and collisions. In simulations based on real e-commerce warehouse layouts, the hybrid approach combining deep reinforcement learning with a planning algorithm achieved approximately 25% higher throughput compared to existing methods and can adapt to different warehouse configurations and robot quantities.

Robohub4/20/2026

AGIBOT introduces Genie Sim 3.0 simulation platform for embodied AI

AGIBOT unveiled Genie Sim 3.0, a simulation platform designed to accelerate embodied AI development by addressing key bottlenecks in environment generation, data scalability, and standardized evaluation. The platform features a spatial world model for generating interactive 3D environments from text/images, a comprehensive benchmark for evaluating five core robot capabilities, and deep integration with reinforcement learning frameworks to enable sim-to-real transfer at scale.

Robotics Business Review4/8/2026

Generalist introduces GEN-1 general-purpose model for physical AI

Generalist AI announced GEN-1, a general-purpose foundation model for robotics that achieves 99% success rates on tasks where previous models reached 64%, while completing tasks three times faster using only one hour of robot data. The model scales from the company's earlier GEN-0 release and uses real-world human movement data from wearables for pretraining rather than expensive teleoperation datasets, with advances in post-training, reinforcement learning, and inference techniques enabling faster adaptation to new robot embodiments and tasks.

Robotics Business Review4/3/2026

Systematic debugging for AI agents: Introducing the AgentRx framework

The article introduces AgentRx, a framework for systematic debugging of AI agents. It appears to be part of a series of developments in AI agent technology, including reinforcement learning integration, human-centered web agents, and autonomous debugging environments.

Microsoft Research3/12/2026

PlugMem: Transforming raw agent interactions into reusable knowledge

PlugMem is a system for converting raw agent interactions into reusable knowledge for AI agents. The article appears to reference multiple AI agent research developments, including work on reinforcement learning integration, adversarial testing for code generation, and external knowledge integration with large language models.

Microsoft Research3/10/2026

The Reinforcement Learning Breakthrough: Why 2025 Is the Year Robots Finally Learn Like Humans

After years of simulation-only training, robotics is experiencing a quiet revolution in real-world learning. New hybrid approaches that bridge low-fidelity simulation with physical reinforcement learning are finally making it practical for robots to acquire complex skills through interaction with their environment—a capability that could fundamentally reshape how we deploy autonomous systems.

Creative Robotics2/12/2026Original

Phi-4-reasoning-vision and the lessons of training a multimodal reasoning model

This article discusses Phi-4-reasoning-vision, a multimodal reasoning model, and explores lessons learned from its training. The piece covers related advances in multimodal AI including reinforcement learning approaches for agents, multimodal reasoning over video and image data, and techniques for improving reasoning capabilities in both small and large language models.

Microsoft Research3/4/2026

How can robots acquire skills through interactions with the physical world? An interview with Jiaheng Hu

An interview with Jiaheng Hu discusses SLAC (Simulation-Pretrained Latent Action Space), a method that enables real-world reinforcement learning for complex robotic systems like mobile manipulators by first learning in low-fidelity simulation then refining in the physical world. The approach addresses key challenges in robot learning including sample inefficiency and safety concerns that have previously limited real-world RL deployment.

Robohub2/12/2026

UniRG: Scaling medical imaging report generation with multimodal reinforcement learning

Researchers have developed UniRG, a system that scales medical imaging report generation using multimodal reinforcement learning to improve the accuracy and efficiency of automated diagnostic reporting. The work demonstrates advances in applying reinforcement learning and multimodal AI to the healthcare domain, combining computer vision and natural language processing for clinical applications.

Microsoft Research1/27/2026

Microsoft Research reveals Rho-alpha vision-language-action model for robots

Microsoft Research announced Rho-alpha (ρα), a vision-language-action model derived from its Phi series that enables robots to perceive, reason, and act autonomously by integrating tactile sensing and learning from human feedback. The model translates natural language commands into control signals for bimanual manipulation tasks and is being evaluated on dual-arm cobots and humanoid robots, with training enhanced through simulation, physical demonstrations, and reinforcement learning.

The Robot Report1/21/2026

Multimodal reinforcement learning with agentic verifier for AI agents

Researchers have developed a multimodal reinforcement learning approach with an agentic verifier for AI agents, with application to medical imaging report generation. The work demonstrates scaling methods that combine reinforcement learning with multimodal inputs to improve automated report generation systems.

Microsoft Research1/20/2026

Customizing multiturn AI agents with reinforcement learning

Amazon Web Services' AI Labs demonstrates that reinforcement learning can effectively customize general-purpose AI agents for specific domains and business use cases, such as customer service and coding assistants, achieving improved task success rates even with small training datasets. The research leverages existing environment simulators and verifiable reward functions to adapt agents without requiring extensive ML expertise or computational resources.

Amazon Science1/13/2026

Did that lamp just fold the laundry? Alumni rethink home robotics

University of Toronto PhD graduates Aaron Tan and Angus Fung co-founded Syncere, a home robotics startup that developed Lume, a robotic floor lamp disguised as a household appliance that folds laundry. The startup shifted from building traditional humanoid robots to creating furniture-integrated robots after learning that customers found industrial-looking robots too intrusive for home environments, using imitation and reinforcement learning with safety features to make the technology practical for domestic use.

Phys.org Robotics1/10/2026

The unseen work of building reliable AI agents

Amazon's AGI Lab is developing reinforcement learning "gyms" to train AI agents on the fundamental micro-interactions required for reliable autonomous task execution, such as scrolling, clicking, form completion, and error recovery across legacy systems. Rather than focusing on high-level tasks like booking vacations, researchers emphasize that agents must master hundreds of low-level atomic behaviors with deterministic reliability before complex real-world applications become feasible.

Amazon Science1/7/2026

Robohub highlights 2025

Robohub's 2025 year-in-review highlights major contributions to robotics and AI research, featuring interviews and articles on robot learning from language instructions, dexterous grasping, human-robot interaction, reinforcement learning for normative behavior, and multi-agent systems. The compilation showcases work from leading researchers and competitions including CoRL, IJCAI, RoboCup, and academic institutions, reflecting significant advances in embodied AI, social robotics, and robot learning methodologies.

Robohub12/29/2025

Continuously hardening ChatGPT Atlas against prompt injection

OpenAI is hardening ChatGPT Atlas against prompt injection attacks through automated red teaming powered by reinforcement learning. The company recently shipped a security update including an adversarially trained model and strengthened safeguards, and describes a rapid response loop designed to discover novel attack strategies internally before they appear in the wild.

OpenAI Blog12/22/2025

LWiAI Podcast #228 - GPT 5.2, Scaling Agents, Weird Generalization

This podcast episode covers major AI developments including OpenAI's GPT-5.2 model with improved multi-modal capabilities, Disney's $1 billion investment in OpenAI for character generation, and DeepMind's research on scaling multi-agent systems. The episode also discusses emerging challenges in LLM factuality, reasoning, reinforcement learning, and AI safety concerns including model corruption vulnerabilities.

Last Week in AI12/17/2025

AI-powered robotic hands learn dexterity by mimicking human movements and anatomy

Researchers at ETH Zurich's Soft Robotics Lab have developed AI-powered robotic hands with 21 degrees of freedom that learn dexterity through imitation learning and reinforcement learning. The hands use tendon-driven artificial actuators inspired by human anatomy and are trained using motion capture data and transformer models to perform complex, unstructured tasks like grasping and sorting objects of varying shapes and sizes.

Phys.org Robotics12/17/2025

Agent Lightning: Adding reinforcement learning to AI agents without code rewrites

Agent Lightning introduces a method for integrating reinforcement learning into AI agents without requiring code rewrites, using multimodal reinforcement learning and an agentic verifier approach. This advancement simplifies the process of enhancing AI agent capabilities through reinforcement learning, reducing implementation friction for developers.

Microsoft Research12/11/2025

Ten years

OpenAI reflects on its ten-year journey since its founding announcement in December 2015, highlighting major milestones including breakthroughs in reinforcement learning, the unsupervised sentiment neuron, and the launch of ChatGPT and GPT-4 that brought AI to mainstream adoption. CEO Sam Altman discusses the company's evolution from a small research team to a major AI company, their commitment to AI safety, and their strategy of iterative deployment to ensure AGI benefits humanity.

OpenAI Blog12/11/2025

GigaTIME: Scaling tumor microenvironment modeling using virtual population generated by multimodal AI

GigaTIME is a multimodal AI system for scaling tumor microenvironment modeling through virtual population generation, while UniRG applies multimodal reinforcement learning to medical imaging report generation. Both represent advances in applying AI and machine learning techniques to healthcare and medical imaging applications.

Microsoft Research12/9/2025

Helping Robots Think on Their Feet

CMU and Meta researchers have developed BFM-Zero, a behavior foundation model that enables humanoid robots to learn and adapt across multiple skills using a single neural policy and unsupervised reinforcement learning. The system allows robots to seamlessly transition between different behaviors like walking, running, fall recovery, and dancing without retraining, moving the field closer to general-purpose humanoids capable of safely operating in real-world environments.

CMU Robotics Institute12/9/2025

Robot Talk Episode 136 - Making driverless vehicles smarter, with Shimon Whiteson

Robot Talk Episode 136 features an interview with Shimon Whiteson, a Professor at Oxford and Senior Staff Research Scientist at Waymo, discussing machine learning and deep reinforcement learning approaches for making autonomous vehicles smarter. The episode explores how techniques like imitation learning are being applied to improve driverless vehicle capabilities.

Robohub12/5/2025

TARS-Like Robot Both Rolls, and Walks

Researchers Aditya Sripada and Abhishek Warrier developed TARS3D, a robot inspired by the character from Interstellar that can both walk and roll using an unconventional four-column design with a shared central hip joint. The project demonstrates how non-biologically-inspired morphologies combined with machine learning and reinforcement learning can enable effective and unusual locomotion gaits.

Hackaday Robotics12/2/2025

How Amazon uses AI agents to anticipate and counter cyber threats

Amazon has developed Autonomous Threat Analysis (ATA), a security system using agentic AI and adversarial multi-agent reinforcement learning to automatically detect and counter cyber threats at machine speed. The system uses red-team and blue-team AI agents that execute real commands on isolated test infrastructure to validate threat detection capabilities and generate improved security rules, reducing typical weeks-long security testing cycles down to hours.

Amazon Science11/24/2025

RL without TD learning

In this post, I'll introduce a reinforcement learning (RL) algorithm based on an "alternative" paradigm: divide and conquer. Unlike traditional methods, this algorithm is not based on temporal differe...

Berkeley AI Research11/1/2025

MIT Researchers Unveil “SEAL”: A New Step Towards Self-Improving AI

MIT introduces SEAL, a framework enabling large language models to self-edit and update their weights via reinforcement learning. MIT Researchers Unveil “SEAL”: A New Step Towards Self-Improving AI first appeared on Synced.

Synced AI6/16/2025

Addendum to o3 and o4-mini system card: Codex

Codex is a cloud-based coding agent. Codex is powered by codex-1, a version of OpenAI o3 optimized for software engineering. codex-1 was trained using reinforcement learning on real-world coding tasks...

OpenAI Blog5/16/2025

A better training method for reinforcement learning with human feedback

Contrasting training pairs with large reward differences mitigate spurious correlations and improve performance of direct-alignment algorithms by as much as 20%–40%.

Amazon Science5/2/2025

DeepSeek Unveils DeepSeek-Prover-V2: Advancing Neural Theorem Proving with Recursive Proof Search and a New Benchmark

DeepSeek AI releases DeepSeek-Prover-V2, an open-source LLM for Lean 4 theorem proving. It uses recursive proof search with DeepSeek-V3 for training data and reinforcement learning, achieving top results on MiniF2F. DeepSeek Unveils DeepSeek-Prover-V2: Advancing Neural Theorem Proving with Recursive Proof Search and a New Benchmark first appeared on Synced.

Synced AI4/30/2025

Scaling Up Reinforcement Learning for Traffic Smoothing: A 100-AV Highway Deployment