Stretch Community News - February 2025
Happy Lunar New Year from Hello Robot!
Recent months saw the release of Stretch AI, a new open-source framework for exploring embodied AI tasks using Stretch containing new tutorials and reference code. Hello Robot also announced a new investment and partnership with the ALS Association, working to evaluate and improve assistive mobile manipulation for people living with ALS.
The Stretch community has also been busy - this month we have a new community-driven Gazebo simulation and tons of new research in assistive robotics, robot learning, open-vocabulary mobile manipulation, and more! Read on for details. If you’d like to see your work featured in a future newsletter, please let us know!
Researchers at Cardiff University have migrated Stretch into Gazebo Sim Harmonic, and released their open-source implementation! Check out their repo below, and watch their presentation from the latest Gazebo Community meeting.
In case you missed it - Stretch AI has been released by Hello Robot! This new open-source set of tools, tutorials, and reference code is designed to help researchers explore the cutting edge of Embodied AI with Stretch, including open-vocabulary mobile manipulation and learning from demonstration! Try it out today with your own Stretch 3 robot.
One common challenge of open-vocabulary mobile manipulation tasks is understanding and adapting to non-static real world environments. DynaMem is a new approach using dynamic spatio-semantic memory to store and update a dynamic semantic map of the robot’s environment, achieving a 2x improvement on state-of-the-art static systems.
RONAR is an LLM-based framework from the University of Washington for generating natural language narrations from robot experiences, combining environmental observations, the robot’s internal state, task planning, and conditions and specifications. The generated narration can be used in downstream tasks like failure analysis, risk estimation, and human interaction and intervention.
Robots generally lack awareness of the noise they produce and its impact on people around them. CMU researchers released ANAVI - a framework enabling home robots to estimate the noise levels in indoor environments using only visual information, and thereby adapt their actions to the noise constraints of the environment.
New University of Michigan work explores how robots and humans can work together to move objects without explicit communication, by embedding subtle signals into the robot’s actions. Their framework uses probability-based inference to predict joint movement strategies, helping the robot minimize uncertainty and improve efficiency, ultimately making it a more effective and natural partner in collaborative tasks.
The newest Stretch Software drop contains information on generating URDFs on-the-fly, using Systemd Unit files to launch applications on boot, a Nav2 bugfix, and a new approach to running Stretch FUNMAP within a UV Virtual Environment.
