SimForcing proposes a simulation-guided framework that injects motion knowledge from a simulation teacher into an action-conditioned robot world model through latent-variable distillation, mitigating appearance differences and reducing the extent to which inaccurate simulation predictions mislead real-video generation.
Sanctuary AI proposes TIGER (Task-Space Imitation Guidance for Efficient Reinforcement Learning), which does not execute the imitation policy but instead maps its action chunks through the controller into end-effector reference trajectories, converting them into dense rewards for reinforcement learning.
Why it matters: The original text provides the specific approach for converting imitation policies into dense rewards along with ablation data, allowing readers to judge whether it can be applied to their own sparse-reward manipulation training.
Michelle Lu, co-founder and CEO of Vsim Technology, discusses using simulation and artificial intelligence to teach robots new skills on the podcast. She previously served as Director of Simulation Technology at NVIDIA, where she and her co-founder developed Isaac Gym, the first end-to-end GPU-accelerated reinforcement learning framework. In 2022, the two left NVIDIA to found Vsim, aiming to address limitations in efficiency, scalability, and the sim-to-real gap in simulation-based robot AI.
At Samsung AI Forum 2026, held at its Seoul headquarters, Samsung Electronics unveiled a robot commercialization roadmap led by its newly established RX (Robot eXperience) Business Promotion Office, combining robot foundation model (RFM) and world model technology to first advance automation at its own more than 100 factories and more than 100 production lines, before entering the home assistant robot market.
Why it matters: The article presents the per-unit cost breakdown of humanoid robots and the success-rate threshold for factory scenarios, which can be used to compare and understand differences in manufacturers' commercialization paths.
At the Samsung AI Forum 2026, Samsung announced its robot commercialization roadmap. Centered on its newly established RX (Robot eXperience) business promotion office, the plan is to first advance automation at its more than 100 factories and more than 100 production lines, then enter the home assistant robot market.
Why it matters: The article lays out the cost math for a humanoid robot at about $100,000 per unit and a three-layer behavior stack, helping readers understand why Samsung puts factory automation ahead of home robots.
The paper LATENT, with Galbot as a participant, received the IROS 2026 Best Entertainment and Amusement Paper Award. The work starts from imperfect human motion data and achieves real humanoid robot tennis rallying. The post provides the project page http://zzk273.github.io/LATENT/, the arXiv address http://arxiv.org/abs/2603.12686, and the code repository http://github.com/GalaxyGeneralRobotics/LATENT, and congratulates the co-authors and collaborators.
Boston Dynamics has unveiled a dexterous hand for its next-generation Atlas humanoid robot, increasing from GR2's 7 degrees of freedom to 13, using a four-finger design with a 4-DOF opposable thumb, direct drive, and only a single actuator type. The design drops the pinky finger to reduce cost and complexity, is equipped with fingertip and palm pressure tactile sensors, supports pinch grasping, three-finger grasping, and pressing tool triggers, and was optimized for high-fidelity simulation for sim-to-real transfer and reinforcement learning, with preliminary sim-to-real results on dynamic tasks already available.
Why it matters: The original article details the degrees of freedom, actuation, and simulation design tradeoffs of the next-generation Atlas dexterous hand, letting readers understand how the hand design serves reinforcement learning and real-world work.
At IROS 2026, Ken Goldberg and Andrea Bajcsy moderated a debate on whether robots should be general-purpose or specialized. Matt Mason opened with "Should animals become human?", and several researchers discussed topics including the gap between tool use and precise manipulation in general-purpose models (coarse placement 19/20 vs. precise insertion 2/20) and the idea that generalists write curricula while specialists execute them. Ken Goldberg called it a "great summary."
Researchers at Georgia Tech have proposed an evaluation benchmark called FindingDory, using 60 tasks that require sustained interaction and awareness of environmental context to test the memory capabilities of vision-language models such as Gemini 2.0-flash and GPT-4o; the tasks cover three categories of high-difficulty scenarios: spatial, temporal, and multi-objective.
At IROS 2026, Zhijian Dynamics debuted a real-robot demonstration of its self-developed world action model SimpleWAM, with the i7 Pro running on-device to complete tasks such as tape sealing and box packing; the i7 Pro and iX demonstrated motion control capabilities together on site.
Why it matters: The original article provides specific optimization data for on-device inference latency and real-robot demonstration scenarios, which can be used to assess the practical level of this model's deployment.
At the IROS 2026 Tech Talk, Zhijian Dynamics presented its precision manipulation work for robot loading and unloading of CNC lathes, where the clearance between the workpiece and the chuck is only 0.5 millimeters; the approach covers motion generation, multi-view perception, historical memory, and contact feedback.
Stial has launched an AI flexible grinding solution for grinding weld seams on welded boxes, achieving uniform removal of weld seams, smooth transitions between welds and base material, and no damage to thin-walled base material. The solution uses the self-developed SeeV vision system to locate weld seams, and NextBrain™ AI Mushroom Cloud computing to calculate grinding allowance and generate grinding paths. After digital twin simulation verification, a six-axis robot works with the PolishX flexible force control system to execute the grinding, with force control accuracy of ±0.5%FS and output force as low as 4N, suitable for energy storage, electrical, new energy battery housings, and other multi-category box scenarios.
KAIST's quadruped robot RAIBO2 finished the Persimmon Marathon in Sangju, South Korea, in 4 hours 19 minutes 52 seconds, slightly faster than the human average, without swapping or recharging its battery, and still had 34% charge at the finish; the team described the work in a Nature paper last week.
Why it matters: The research moves energy saving for legged robots from optimizing single components to holistic design across joints and circuits, offering concrete methods transferable to long-duration outdoor robots.
At IROS 2026, under the theme "From Touch to Intelligence," Daimon Robotics publicly demonstrated its tactile intelligence infrastructure on real hardware for the first time, covering three layers: the DM-Tac series of vision-based tactile sensors (more than 12 tactile modalities and over 110,000 sensing units), the Data-Nexus data collection system, and the Daimon-TWM world model released in August.
We are releasing Dyna-2.1, the first physical agent to achieve reliable super long-horizon whole-body autonomy. It combines our new semi-humanoid hardware with Da...
Being-H0.7, the general-purpose dexterous manipulation large model independently developed by ZhiZaiWuJie, has been accepted to NeurIPS 2026, marking the Being-H series' completion of acceptances to the three top conferences ICML, CoRL, and NeurIPS within one year.
Simplexity Robotics has launched and open-sourced the world-action model SimpleWAM; the 6B-parameter version reduces inference latency on NVIDIA ThorX (T5000) from 181.23ms to 42.31ms, enabling high-frequency closed-loop operation on-device at nearly 24Hz.
Why it matters: The original article presents the on-device optimization path from 181ms to 42ms and four architectural designs, which can be transferred to understand how world-action models can meet high-frequency closed-loop control requirements.
Lingchu Intelligence has officially released its embodied model Psi-R2.5, which uses a two-tier architecture with a QwenVL3.5-4B high-level planner plus a Wan2.2-IT2V-5B low-level controller, and proposes the strong Pair Data standard. The model reverse-invokes the Psi-W0 world model to generate human demonstration videos backward from real-robot data, enabling mass production of paired data, and supports in-context learning (ICL) for zero-shot generalization to new tasks. It comes with a HIL+RL post-training framework and is claimed to raise task success rates to 99% after a few iterations taking only 1-2 working days, and has already been deployed at customer sites.
Why it matters: The original article details the model architecture, the strong Pair Data data standard, and the reverse-generation method, letting readers understand how its data approach supports real-robot generalization.
Professor Zhang Tao, Dean of Tsinghua University's Institute for Embodied Intelligence and Robotics, delivered a lecture titled "Embodied Intelligence Empowers Everything" at the Tsinghua Humanities Forum. He described three major characteristics of embodied intelligence—generality, interactivity, and autonomy—and noted that key technologies still need to be breakthroughs, including multimodal embodied large models, environmental perception, dynamic planning and control, and dexterous bodies.
Our robots are not driven by a single AI model. This is a deliberate choice. The first installment in our 10-part Physical AI series: where AI holds up in production, where it does not, and why safety sits outside the learning loop. https://www.atirobotics.ai/resources/educational-content/where-we-stand-on-ai/
Mondo Robotics announced that the open-source reinforcement learning environment for Beni is now live, allowing developers to train their own policies with MJLab and test them on the track. The environment already provides entry points on Hugging Face Space and GitHub. The quoted content states that Beni is the most-funded personal robot on Kickstarter.
Self-play is a powerful path for physical AI. @SkildAI's model trains against increasingly stronger versions of itself, turning a simple objective into emergent behaviors.
Noetix released the technical report for GLOW (Generative Learning of World), a general embodied intelligence generative learning framework, claiming an average success rate of 62.2% across 18 tasks on RoboDojo.
A team at Carnegie Mellon University's Robotics Institute has developed the LAMP (Long-Horizon Adaptive Manipulation Planning) system, which helps multiple robots collaborate to move objects in severely cluttered spaces by jointly planning object and robot motion rather than handling them separately.
A KAIST team published RAIBO2, a quadruped robot, in Nature; it completed the Sangju Marathon in South Korea on a single charge in 4 hours and 19 minutes, running alongside human runners.
ETH Zurich's Soft Robotics Lab taught an off-the-shelf humanoid robotic hand to walk autonomously, press arrow keys, and push blocks. With a battery, motion sensors, and a microcomputer added to the back of the hand, and after reinforcement learning training in simulation, the hand crawled on 14 indoor and outdoor surfaces, and recovered its standing posture on its own in 21 of 25 sideways-fall tests. It pressed the correct arrow key 29 out of 32 times. In 15 tests, the block was pushed to within 2 centimeters of the target. The paper Fingers as Legs has been published on arXiv.
A new blog post co-written with @perryadong discussing what robots need to be broadly useful in the real world. https://pd-perry.github.io/posts/post-training.html Reliability is one of the biggest unsolved challenges in AI today. Current models do okay when a human reviews their output (such as drafting code), but this will become a bigger bottleneck when we want systems to operate with greater autonomy and greater trust.
Real-Time EXPO-FT enables π0.5 to tackle challenging dynamic tasks such as balancing a ball and striking a ball into a goal; the paper page is https://pd-perry.github.io/real-time-expo-ft/. Chelsea Finn explains that the approach handles latency in reinforcement learning by having the base VLA use older images while a small editing policy runs in real time, and states that it significantly outperforms RTC across comparisons.
Zero Power Robotics appeared at the 2026 CIFTIS as a representative company expanding embodied AI overseas; its robot convenience store project demonstrated an unmanned retail loop in which, after a customer scans a code to place an order, the robot autonomously completes product recognition, grasping, and delivery. Recently, multiple Zerith H1 wheeled-arm humanoid robots have been deployed at a laboratory in Malaysia for local solution providers to conduct secondary development, data collection, and model training. The company is continuously iterating its Zerith V series embodied-native motion models and advancing the API-ization of embodied AI physical grasping.
Skild AI announced that it post-trained its foundation model S1 through physical self-play, letting the policy play against a recent version of itself with scoring as its sole objective. In NVIDIA Isaac Sim, over 140 years of simulated time, it learned skills such as dribbling past opponents, shielding the ball, and tackling, and these skills were successfully transferred to humanoid robots for matches. The authors believe self-play can break through the capability ceiling of pretrained human data, and the next step will be to extend it to everyday tasks such as collaborative manipulation and city-scale navigation.
Why it matters: The authors describe using physical self-play to let a foundation model acquire soccer skills in simulation and transfer them to real robots, which lets readers learn about a feasible path for self-play-style post-training.
MagicLab introduces the visual guidance approach in the Magic-Brain framework: a VLM Agent outputs a bbox, and a segmentation model refines it into a three-color mask overlaid on the observation image, converting semantic guidance into visual guidance.
Why it matters: MagicLab details the design motivation and implementation of Magic-Brain's visual guidance layer, letting readers understand the impact on grasp success rate of converting language instructions into pixel guidance.
Sergey Levine shared research conducted with Siemens and Berkeley on how to pair a small RL policy with a large robot foundation model under real-time chunking (RTC). Because the RL policy infers faster, it can observe newer images, thereby steering the foundation model's policy toward better behavior. The introduction he cited notes that asynchronous VLA inference reduces inference latency but breaks the Markov assumption required for reinforcement learning fine-tuning, and ARLI is proposed to address this problem. The project page is https://async-rl-intermediate-information.github.io/ . The work was led by Brian Zhu, Momen Khalil, Emanuele Poggi (Siemens), and @ehharrison4 (Berkeley), among others.
Ruoyu Technology was invited to attend the 2026 Energy Industry Science and Technology Innovation Integration Development Conference, where founder and CEO Sun Teng delivered a keynote speech introducing the self-developed Ruoyu Jiutian robot brain. This robot brain features multimodal perception, long-sequence task planning, autonomous decision-making and execution, and other capabilities, providing underlying intelligent support for robot bodies; based on the embodied intelligence brain system solution for high-risk energy scenarios built on it, the special explosion-proof robot Ruoyu Lanyue was launched. Following its debut at the Chain Expo, Ruoyu Lanyue recently entered a regular operating gas station, launching pilot operation of multiple new functions.
The Embodied Intelligence and Intelligent Computing Development Forum of the 2026 AI Computing Conference (2026AICC) was held in Beijing on September 21, hosted by Tsinghua University's Institute for Embodied Intelligence and Robotics, under the theme "Breaking Through the Critical Point of Human-Machine Symbiosis." Scholars including Zhang Tao and Zhao Mingguo of Tsinghua University, along with representatives from companies such as Lingyu Intelligence, Shouyi Technology, and Qianjue Technology, shared results on directions including clinical applications of embodied intelligence, brain-inspired approaches for humanoid robots, data infrastructure, and world models. The forum built an exchange platform for industry, academia, and research, promoting the engineering and deployment of embodied intelligence technologies.
Dexbotic has fully integrated DM0.5, upgrading both its end-to-end pipeline capabilities and its evaluation coverage. Readers can use "Read the original article" to go directly to the GitHub repository and view past issues.
At the Greater Bay Area Film Concert, AGIBOT used the WITA-Omni omni-modal embodied model to achieve real-machine human-robot interaction; the robot can identify its conversation partner, judge the right timing to respond, and coordinate speech, motion, and facial expressions in real time. The model uses a Thinker-Talker-Actor architecture, and WITA-Omni Preview ranked first on the DailyOmni benchmark with an average accuracy of 85.21%, leading in six of eight evaluation dimensions.
Psi Robotics released the Psi-R25 robot manipulation model. Its core upgrades are improved data quality and data attribute compression: it reviewed and extracted high-quality data from 100,000 hours of data, and used strong pair data to align human data with robot data.
Why it matters: The original article details how strong pair data is generated and the data quality standards, giving readers a referenceable path for aligning human data to robot data.
On September 18, at the EBC 2026 Embodied Intelligence and Humanoid Robot Brain Technology Conference in Shanghai, RoboCT Robot received the 2026 China Embodied Intelligence "Yingxing Benchmark Enterprise Award." At the conference, software algorithm engineer Liu Xiaopeng delivered a keynote speech titled "Engineering Practice of VLA Models in Complex Assembly for Humanoid Robots." Its second-generation Hongjun wheeled dual-arm robot has entered small-batch pilot procurement and completed deployments across multiple industries, and the YiBrain large model has been validated on real production lines.
Ken Goldberg of UC Berkeley argues in an article for the Agentic Robotics (AR) approach, in which multi-agent systems write, test, and iterate structured robot control programs offline, as a third path between model engineering and data-driven methods that does not rely on large-scale robot demonstration data.
A third-party technical report, "GPT-6 Astra as an Embodied Policy," co-authored by Galbot researchers, was open-sourced less than two weeks after GPT-6 Astra's release, testing its performance as an embodied policy in the RoboDojo and RoboLab simulation environments.