Hello from all of us at Hello Robot!
This edition of Stretch Community News showcases impressive progress across the robotics landscape. You’ll find research on autonomous manipulation that learns from human video and self-corrections, lightweight safety-critical control frameworks for foundation models (FEARL), long-horizon decision-making under uncertainty (ROP-RAS3), and geometry-aware vision-language navigation (GA-VLN). We’ve also included a clever project from Arizona State University where a Stretch 3 uses reinforcement learning and LLMs to master the Three-Card Deception trick.
Read on for more details! And if you’d like your work featured in a future newsletter, we’d love to hear from you. Drop us a note at community@hello-robot.com.
Cheers,
Aaron Edsinger
CEO – Hello Robot
Hello Robot has open-sourced a new autodocking system for Stretch 4, bringing the robot one step closer to fully unsupervised operation. The system uses LiDAR-based dock detection, local obstacle mapping, and sampling-based control to help Stretch safely navigate to and connect with its charging station. With low-latency control and obstacle avoidance, the system enables reliable docking across different environments and floor types.
We were recently featured in TechCrunch for our work bringing practical, AI-powered robots into everyday environments. The article highlights our approach of building robots for real-world use today, helping researchers and advancing assistive robotics. It’s exciting to see our mission recognized as we continue working toward a future where robots can make a meaningful difference in people’s daily lives.
Researchers from ETH Zürich, the Technical University of Munich (TUM), and Microsoft introduced a new approach that helps robots improve their own performance by learning from human videos and their own mistakes. Rather than relying on constant human corrections, the system uses AI models to predict better actions after failures, significantly improving success rates across a variety of real-world manipulation tasks. The work points toward more autonomous robots that can continuously refine their skills with minimal human intervention.
A new framework from University of California, Irvine combines powerful AI foundation models with a lightweight, verifiable safety system for robot control. Called FEARL, the approach separates high-level planning from safety-critical actions, allowing robots to take advantage of advanced AI while providing formal guarantees for behaviors like collision avoidance. The research offers a promising path toward making autonomous robots both more capable and more trustworthy.
A collaboration between the Australian National University and Rice University introduced ROP-RAS3, a new planning algorithm that helps robots make better decisions over long time horizons in uncertain environments. By rapidly sampling possible actions instead of evaluating every option, the approach significantly outperformed existing planning methods across a variety of complex tasks and was successfully demonstrated on a physical robot. This work could enable more capable autonomous robots operating in dynamic, real-world settings.
A team from Arizona State University developed a new reinforcement learning framework that enables robots to autonomously solve the classic Three-Card Deception trick using a Hello Robot Stretch 3. By combining strategy discovery, computer vision, and logical reasoning, the system can infer a participant’s hidden card choice without relying on predefined heuristics. The work demonstrates how reinforcement learning and large language models can create more capable and engaging human-robot interactions.
A team from the Chinese Academy of Sciences, University of Chinese Academy of Sciences, National University of Singapore, and The Hong Kong University of Science and Technology, along with researchers from Robbyant, developed GA-VLN, a new approach to vision-language navigation. The system creates a compact, 3D-aware bird’s-eye-view representation that helps robots better understand their surroundings while reducing the amount of visual information an AI model needs to process. GA-VLN achieved state-of-the-art navigation performance while using less data and computational resources, and was also demonstrated on a Hello Robot Stretch 3 in a real-world environment.
We’ve introduced a new lightweight data collection workflow for Stretch 3 designed to make imitation learning easier. The recorder runs directly on the robot without ROS 2, capturing synchronized robot telemetry, teleoperation actions, and camera imagery, which can then be converted into a LeRobot dataset on a separate development machine. This provides a simpler path from teleoperating Stretch to creating high-quality datasets for training and advancing robotic learning.
In a recent piece, Zen KOH explores what makes a robot genuinely useful in the real world through a conversation with Hello Robot CEO Aaron Edsinger. The article contrasts Stretch’s intentionally non-humanoid design with the current push toward humanoid robots, highlighting how usefulness, autonomy, and human-in-the-loop design can matter more than mimicking human form. It offers a thoughtful look at why building robots around real human needs, not simply human anatomy, could be key to making robotics truly useful in everyday life.