Stretch Community News - September 2025
Hello again from Hello Robot!
What an exciting couple months it has been! We were honored to demo Stretch 3 at the AI for Good Summit in Geneva, where over 17,000 participants and 800 speakers gathered to help move from AI theory to practice. The event showcased AI’s potential for advancing the UN Sustainable Development Goals, with new initiatives in food systems, multimedia authenticity, healthcare, climate solutions, and global standards.
Back in the Stretch community, we’ve seen some exciting new research happening: studies on how human feedback strategies shape robot learning, long-term integration of care robots in senior living facilities, and Princeton’s “invisible” delivery robot for VR/AR. Other highlights include Fail2Progress for learning from manipulation failures, the FINDINGDORY benchmark for memory in embodied agents, eFlesh magnetic touch sensing that improves manipulation success by 40%, and some promising evaluations of robot-led exercise for patients with Parkinson’s disease.
Lots to celebrate this month, read on for details! And if you’d like your work featured in a future newsletter, we’d love to hear from you.
This study examined how human feedback strategies affect machine learning. Thirty-six participants gave numeric feedback to a robot during a card game, with varying “partial credit” for the same errors. When used to train models, some feedback strategies produced much stronger learning outcomes than others. Interestingly, participants’ background knowledge did not significantly impact results, highlighting the importance of human teaching methods in real-world ML.
This dissertation explored how robots can be integrated into senior living facilities to address caregiver shortages. Rather than focusing on short-term interactions, it emphasized long-term integration into daily workflows and relationships. The author developed CareAssist, an end-user development tool, showing that empowering caregivers and older adults to adapt robots improves adoption. End-user development is highlighted as a key part of building safe, effective care robots.
Princeton researchers built an “invisible” mobile robot system that delivers objects to VR users without breaking immersion. Using a Quest 3 headset, users see virtual items while the robot moves the real object into place. The system can even disguise the robot in VR, blending physical and digital experiences for seamless interaction.
The authors propose Fail2Progress, a method for helping robots learn from failures in long-horizon manipulation tasks. Using Stein variational inference, the system generates parallel simulations of failure cases to create efficient training data. Tests on tasks like shelf organization and object transport showed that Fail2Progress enabled better recovery and outperformed baseline methods.
This work examined large vision-language models (VLMs) for robotics, noting their struggles with long-term memory across many images. To address this, the authors created a new benchmark in the Habitat simulator with 60 tasks requiring navigation, manipulation, and contextual awareness. Baseline tests showed current VLMs fall short in memory-intensive scenarios, pointing to areas for improvement in embodied AI.
The eFlesh project introduced a low-cost, customizable tactile sensor for robotic manipulation. Built with a 3D printer, magnets, and a magnetometer board, it allows tuning of geometry and sensitivity through modular microstructures. Experiments showed high accuracy in force prediction and slip detection, while integration into control policies improved manipulation success by 40% compared to vision-only systems.
This study evaluated a robot-led rehabilitative exercise system for individuals with Parkinson’s disease. Eleven exercise specialists tested the system and provided feedback through questionnaires and interviews. Results showed positive reception, with potential to improve patient engagement and support therapy, though participants suggested more natural feedback and easier use.
