Magmotor Technologies Inc., founded in 1876, has been manufacturing custom AC and DC fan motors for over 150 years. The company plans a major expansion that will triple its income and number of employees. Magmotor specializes in brush motors, even as many other companies have converted to brushless motors, due to legacy items still needing these motors.
Generative Bionics has launched GENE.01, a multimodal smart-skin humanoid platform that incorporates distributed tactile skin sensing touch, force, temperature, and proximity for safe physical interaction with human workers. Built in six months after an $81 million seed raise, the platform utilizes physics-native AI to co-optimize physical hardware and motor control around human ergonomic safety. The company is open-sourcing the robot's digital twin across PyPI, Conda, and ROS ecosystem repositories.
Highline Commerce has partnered with Ultra Robotics to automate up to 30% of its order volume using humanoid robots at Brooklyn's Industry City tech hub. The robots, which are designed for continuous 24-hour operation without battery swaps or tipping risks, are handling pick, pack, scan, and conveyor processes for over 200 retail brands. Highline Commerce is one of the first fulfillment companies to deploy humanoid-style robots for live, commercial retail operations.
Figure has officially manufactured its 1,000th Figure 03 humanoid robot at its BotQ facility, marking a significant acceleration in production. The milestone comes roughly three months after the company announced scaling its throughput to one robot per hour. This achievement solidifies Figure's position as one of the few Western companies capable of moving beyond bespoke prototyping into legitimate batch manufacturing.
A small Dual physics gate has been built to test robot claims, with a boolean gate coming from Rust. The gate is designed so that the "Safe" path allows a claim while the "Hostile" path refuses it. A production-ready version of the gate has been created and can be cloned and run. It also generates sealed receipts for validation purposes.
An open-source ROS 2 Text-to-Speech (TTS) Plugin for Gazebo Harmonic has been released, supporting Gazebo Classic and modern versions with ROS 2 Humble and Jazzy. The plugin was redesigned from the original Gazebo Classic version and is available on GitHub as gz_ros_tts. It has been successfully tested on various robots and environments, including the Heinz H1 Humanoid Robot.
Researchers have proposed a new framework using Generative AI to create realistic and diverse sequences of human body movements, enabling robots to learn complex skills without real-world data. This approach converts textual prompts into synthetic demonstrations, allowing the robot to observe multiple variations of task-execution styles. The method is evaluated across four simulation scenarios and shows strong adaptability to complex motion variations.
Recent studies have increasingly explored progress rewards in robotic learning to provide feedback during task execution, as terminal success signals only indicate completion but not the nature of behavior. However, existing methods use different observations, goal specifications, output signals, supervision sources, and evaluation protocols, making it difficult to compare them. This survey provides a unified view of progress reward modeling for robotic learning, organizing the field into three connected steps: studying the interface, moving inside the model, and examining data and benchmarks.
GRACE, Gradient-free Robot Action generation via Combined diffusion-MPPI posterior mean Estimation, guides a pretrained diffusion policy with Model Predictive Path Integral (MPPI) control using only forward cost evaluations. GRACE constructs a cost-conditioned guidance posterior at each reverse step and estimates its mean with a single MPPI update centered at the diffusion reverse mean. In simulation, GRACE attains higher success rates than diffusion-based and sampling-based baselines, while on a real 7-DoF manipulator, it avoids a deployment-time obstacle that unguided priors collide with in every trial.
Ordered Action Tokenization (OAT), a learned action tokenizer, discretizes robot action chunks into an ordered sequence of tokens using a transformer with registers, finite scalar quantization, and ordering-inducing training mechanisms. OAT places coarse control information in early tokens and refines residual detail using later tokens, yielding an anytime tradeoff between inference cost and action fidelity. It consistently delivers strong policy performance while offering significantly greater flexibility at inference time across three policy backbones and over 60 tasks spanning five simulation benchmarks and real-world settings.
Researchers Liane Galanti, Dhruv Shah, and Tri Dao introduced a new approach to addressing the orchestration gap in generalist robots by creating a Physical Agency orchestrator (Pigey). This closed-loop physical agent can control existing vision-language-action policies and parametrized skills to solve complex reasoning tasks without additional data collection or post-training. Pigey demonstrates significant performance improvements over existing generalist policies, advancing the state-of-the-art on simulation benchmarks and real-world robotic manipulation tasks by over 4x.
StARS, a novel model-agnostic framework, integrates collaborative filtering with learnable scene representations to generate user-specific appropriateness scores over candidate robot actions. It treats annotators as users, contexts/scenes as items, and appropriateness scores as targets. StARS consistently improves performance and agreement with annotators across two socially aware robotics datasets, MannersDB+ and SocNav1. The framework is model-agnostic, enabling personalization without redesigning the underlying model.
Researchers Siqi Fan and colleagues have developed an action-conditioned world model framework for goal plane probe guidance in robotic ultrasound, which achieves success rates of 70.0% for carotid guidance and 65.0% for thyroid guidance in real-world closed-loop experiments. The framework uses a two-stage pipeline that combines latent conditional diffusion and goal-conditioned temporal transformer to predict future ultrasound observations and ordered probe motions. This approach addresses the challenge of collecting high-quality demonstrations for robotic ultrasound tasks due to the labor-intensive nature of training and the difficulty of building explicit simulators.
Deep reinforcement learning enables adaptive locomotion of snake-like robots in dynamically changing viscous environments, overcoming traditional control methods. A teacher policy distills knowledge into a student policy relying solely on proprioceptive sensor information. The DRL agent autonomously acquires non-sinusoidal gaits that improve propulsion velocity and transport efficiency across a wide range of dynamic viscosity changes ($10^{-7}$ to $10^{-2} m^2/s$).
A large-scale dataset called ACME has been introduced, capturing goal-driven social navigation behavior in complex scenarios with explicit robot-crowd interaction. The dataset was collected across 8 sites in 5 countries using 7 robot embodiments, providing 29.35 hours of onboard robot data and 43.5 hours of overhead pedestrian tracking data. It includes 3D and 2D scene features, odometry, interaction information, and human-annotated pedestrian trajectory labels to facilitate learning navigation policies and predicting pedestrian trajectories. The dataset aims to advance social navigation research by addressing the lack of diversity in existing datasets.
A large-scale outdoor geomagnetic SLAM dataset, Mag4D-SLAM, was created to support research on yaw drift mitigation and location-discriminative magnetic signatures for cross-session place recognition. The dataset comprises 14 sequences totaling over 18 km of synchronized LiDAR, camera, IMU, tri-axis magnetometer, and GNSS measurements with SE(3) ground-truth poses. It was collected along structured campus trajectories under paired day/night conditions in both forward and reverse directions, analyzing three core properties: magnetic field repeatability, drift-free global heading estimation, and location-discriminative magnetic signatures. The dataset is designed to complement visual and LiDAR modalities or provide a fallback cue under illumination changes.