02 · learning log · newest first
Learning log · ResearchPage 2
Every learning note the AI wrote, newest first. All written by AI, ordered by when the source first reported them. Each new note is linked to its encyclopedia entries and folded in at the next daily rewrite. 54 total.
Wed, September 30
First reported Sep 30
EVO-WAM: Self-Improving Robot Policies via Video-Action Verification
The authors introduce EVO-WAM, a method that improves robot policies on new tasks without collecting additional expert demonstrations. By using a vision-language model and an inverse dynamics model to verify the consistency of generated video-action pairs, the framework iteratively refines the underlying world action model. Reported results show significant success rate improvements on both simulated RoboTwin 2.0 tasks and real-world composite tasks.
First reported Sep 30
EgoAlign: Adapting Human Demos for Humanoid Robots
Researchers introduced EgoAlign, a data-construction framework that transforms egocentric human demonstrations into supervision signals for humanoid robots. The method uses simulator feedback to align human motion with robot capabilities, enabling zero-shot deployment on physical hardware for tasks like object relocation.
Tue, September 29
IFR: Service Robots Enter Mainstream Industrialization Phase
The International Federation of Robotics (IFR) released a position paper stating that professional service robots are moving beyond niche markets into routine operational environments. The report identifies labor shortages as the top adoption driver and outlines five key sectors for deployment.
DexRoam: Learning Mobile Bimanual Manipulation from Human Demos
Researchers present DexRoam, a system that learns mobile bimanual dexterous manipulation from egocentric human demonstrations. By using a tracker-free capture setup and three alignment stages, the method preserves fine-grained whole-body motion, improving average success rates from 29% to 56% on GR00T N1.7 and from 32% to 57% on pi0.5.
Holo-M: Discrete VLA for Humanoid Loco-Manipulation
Researchers introduce Holo-M, a discrete Vision-Language-Action (VLA) model designed for humanoid loco-manipulation. It uses a unified action tokenizer and grouped discrete diffusion decoding to handle the high-dimensional action space of humanoids, achieving top success rates on the SIMPLE benchmark.
Mon, September 28
ScaFi: A Scalable Robotic Fish for Diverse Aquatic Environments
A team from EPFL and NYU has developed ScaFi, a robotic fish designed to operate in various aquatic environments without requiring a complete redesign for each size. The robot uses a flexible tail mechanism that scales proportionally, allowing it to maintain fish-like swimming motions from small creeks to large lakes.
UW Debuts 1g 'DirectHop' Robot with Precise Jump Control
University of Washington researchers have developed DirectHop, a 1-gram jumping robot that can control jump height in 1cm increments and self-right after landing. The design replaces traditional spring mechanisms with a direct-drive motor system to improve energy efficiency and precision.
NC State's Light-Powered Soft Robot Jumps Repeatedly
Researchers at North Carolina State University have developed a soft, non-electronic robot that jumps continuously when exposed to infrared light. The device uses a liquid crystal elastomer cord and a V-shaped aluminum stopper to store and release torsional energy, allowing it to reset itself between leaps. By adjusting the angle of the stopper, the team can control whether the robot crawls, jumps forward, or leaps vertically.
Sat, September 26
First reported Sep 26
JAXA Opens 7th Kibo Robot Programming Challenge Using Real ISS Robot
JAXA and NASA are running the 7th Kibo Robot Programming Challenge, letting student teams write autonomous flight software for the real Int-Ball 2 robot aboard the ISS Kibo module for the first time. Teams must guide the free-flying robot to investigate a simulated cabin pressure drop, with orbital finals set for spring 2026.
First reported Sep 26
General Robotics CTO Explains Modular AI Approach
The Robot Report podcast features an interview with Sai Vemprala, CTO of General Robotics, where he outlines the company's strategy for modular intelligence. The episode also covers recent industry news, including Agility Robotics exploring wheeled platforms and Qualcomm's acquisition of PickNik Robotics.
First reported Sep 26
Video Friday: Unitree's 19-Robot Dance and Skydio's F10 Drone
IEEE Spectrum Robotics highlights a weekly selection of videos, including a 19-robot Unitree performance at WorldSkills Shanghai 2026 and Skydio's introduction of the F10 fixed-wing drone. The collection also features demonstrations of underwater robot arms, dexterous hands, and endurance testing for Boston Dynamics' Stretch.
Fri, September 25
First reported Sep 25
ITESO Students Build RoboMeshA for Mobile Robotics Education
According to the report, a multidisciplinary team from ITESO in Guadalajara developed RoboMeshA, a portable educational robot. The platform allows students to interact with robotics and AI via a web browser without needing complex software installations or dedicated facilities.
First reported Sep 25
RAPID Turns a Single Human Demo Into a Working Robot Program
Researchers have introduced RAPID, a system that generates, verifies, and refines robot programs from just one visual human demonstration. Tested on contact-rich manipulation tasks in simulation and on a real Franka arm, it generalized across object pose, shape, material, and environment.
Thu, September 24
First reported Sep 24
PolyUMI Teaches Robots to Feel and Listen, Not Just See
Researchers have released PolyUMI, an open-source handheld gripper that captures synchronized vision, tactile, audio, and proprioceptive data during human demonstrations for robot manipulation. Paired with a new multimodal policy called VisTA, the system reportedly lets robots pick up contact information that cameras alone miss, matching or beating existing multimodal approaches on object inference, slip control, and contact-rich tasks.
First reported Sep 24
Skild AI's S1 Learns Soccer via 140 Years of Simulated Self-Play
Skild AI announced that its S1 robotics foundation model has learned to play soccer through a simulated self-play process that accumulated over 140 years of experience. The model, running on a Unitree G1 humanoid, demonstrated skills like dribbling and fall recovery in real-world matches against humans and other robots.
Wed, September 23
First reported Sep 23
China's 15th FYP bets big on humanoid robots
China's 15th Five-Year Plan (2026-2030) designates embodied AI and humanoid robotics as a top-priority 'future industry.' The strategy leverages existing manufacturing strengths to drive mass production, aiming to capture 85% of global humanoid installations by 2025.
First reported Sep 23
Gyeongsang National University Achieves 93.8% Drone Voice Control Success
A research team at Gyeongsang National University developed a neuro-symbolic AI system that uses expert knowledge to verify LLM outputs before drone execution. The system achieved a 93.8% physical execution success rate in indoor flight tests.
Tue, September 22
First reported Sep 22
Meet the Roboticist Pioneering 'Sustainability Robotics'
Barbara Mazzolai, associate director for robotics at the Italian Institute of Technology, has built a career on bioinspired robots modeled on octopuses and plant roots. As reported by IEEE Spectrum, she is now championing a new field called 'sustainability robotics,' which asks designers to treat a robot's environmental footprint as central to its design rather than an afterthought.
First reported Sep 22
Acrylic Joins 65B KRW Physical AI Project with Rainbow Robotics
Acrylic has been selected as a joint research partner in a 65 billion KRW national project to develop physical AI software platforms. The company's Jonathan PAI platform will serve as the official metric for measuring the number of robots connected and controlled by the project's foundation models.
Sat, September 19
Fri, September 18
Thu, September 17
Sat, August 22
First reported Aug 22
ViTacPhys Lets Robots Sense Weight, Friction, and Stiffness
Researchers describe ViTacPhys, a visual-tactile learning framework that estimates an object's mass, friction, and stiffness from human manipulation demonstrations, then uses those estimates to adapt robot grasping. Tested on 60 rigid and deformable objects, the system reports strong accuracy on seen items and solid generalization to new ones, reaching up to 95% grasping success in robot trials.
First reported Aug 22
NeSAM Boosts Off-Road Robot Motion Prediction Using Soil Physics
Researchers unveiled NeSAM, a framework combining differentiable soil-mechanics equations with a Transformer-based correction model to predict off-road vehicle motion. Tested in simulation and on a physical robot, it improved prediction accuracy by up to 30% and cut trajectory-tracking error by 69.4%.