Stretch Community News - July 2025
Hello again from Hello Robot!
We've been thrilled to see so many of you recently at events like ICRA in Atlanta, CVPR in Nashville, and RSS in Los Angeles. Tons of great work was presented with Stretch, including live research demonstrations by teams from NYU and UIUC. ICRA also saw the exciting on-robot conclusion of the PhyRC Challenge, where the RoboNotts team from the University of Nottingham took home the victory and won their own Stretch 3! Congratulations!
Lots to highlight from the Stretch community - this month we have a new approach to Embodied Question Answering, robot profanity, avocado grasping, a robotic nursing tutor, an exciting new MuJoCo update, and more! Read on for details. If you’d like to see your work featured in a future newsletter, please let us know!
Embodied Question Answering requires robot agents to explore and understand their environment, in order to answer questions like “Is there a blue pan on the stove?”. Researchers from Carnegie Mellon University and the Bosch Center for AI present GraphEQA, a new approach to this problem that builds real-time 3D semantic scene graphs and uses them with vision-language models. This approach enables more efficient exploration and improves question-answering accuracy compared to prior techniques.
Oh &#$*! Prof. Naomi Fitter’s robotics lab at Oregon State University explored how robots using curse words when they make a mistake could affect how people perceive them. Across three experiments, the researchers found that while it's helpful when robots acknowledge errors, adding profanity had some mixed results - some found it disconcerting, but many thought it was humorous and endearing, suggesting room for more playful robot designs.
UIUC researchers explored a new modular system (MOSART) for end-to-end drawer and cabinet opening in novel, real-world environments. After conducting 100+ real-world trials, they found that modular systems outperformed even end-to-end learned approaches with thousands of demonstrations, and noted that perception - rather than precise control - emerged as the main bottleneck.
Information asymmetry is a key challenge in training student policies from teacher policies, like robots learning from human examples. Researchers at Cornell University created two tools to help make this process more effective: CritiQ, an imitation learning approach which lets the student know when to ask for help by predicting states with no recoverable supervision; and ReTRy, a reinforcement learning approach which selects the best state to initialize training for efficient exploration. In simulation and on Stretch, these approaches demonstrated significant improvements over standard teacher-student baselines in training efficiency and final performance.
New work in agricultural robotics from Khalifa University and Abu Dhabi University demonstrates a novel, non-invasive approach to estimating fruit ripeness using Stretch with DIGIT tactile sensors. Their new SwishFormer model, paired with a Random Forest regressor, accurately predicts fruit firmness and ripeness by palpating them, outperforming existing methods and achieving up to 98% success in testing with avocados.
Amid shortages of trained nursing educators, researchers from Rice University and Houston Methodist have developed ASTRID, a robotic tutor designed to help train nurses to perform central line dressing changes, improving prevention of serious life-threatening infections. Built through a two-year participatory design process, ASTRID delivers real-time feedback and guidance, with early evaluations showing strong potential to enhance training and reduce preventable complications.
A new release (v0.5) of Stretch MuJoCo, a high-fidelity simulation of Stretch 3 built on the MuJoCo physics engine, focuses on closing the gap between real and simulated behavior so software written against simulation can translate more readily to a real Stretch. This release adds real-time simulation support, including depth/color cameras and 2D lidar, as well as ROS2 support for applications including Nav2 and Stretch Web Teleop. It can also now generate Robocasa kitchen-style environments, and integrate with other custom user-generated environments. Stretch MuJoCo can run headless in Google Colab, and includes tutorials and docs for getting started alongside an advanced API for features like cloth and deformable simulation and procedural model generation.