Claudia D'Arpino 🤖🧠

The ChatGPT of robotics needs both a model (GPT) AND an interaction method (“chat”) to let users steer robot policies online, post-training. The multimodality in diffusion models is great for capturing diverse behaviors, but users need control at inference time! @felixwyw & team’s work at #ICRA2025 presents a method to steer pre-trained policies during generation. Check out yanweiw.github.io/itps/ for all the results!

🧠Reasoning VLAs generate textual plans 💬 via CoT before predicting actions 🦾. But do the predicted actions actually follow the plans? Introducing SEAL: a training-free runtime steering method that ensures VLAs "do what they say". Website: tinyurl.com/seal-vla (1/9)

Excited to invite you to our #RSS2025 Workshop at RTH 109, where we’ll explore the frontier of Generative Models × Human–Robot Interaction! 🤖✨ Organizing with @robominyoung @sammy_j_c @haroldsoh @andreea7b @julie_a_shah sites.google.com/view/gai-hri/h… 9:00 AM – Workshop Intro • @robominyoung 9:10 AM – Visual Embodied Planning • @RoozbehMottaghi 9:40 AM – Simulating People with Language Models—Good Idea? Bad Idea? • @ybisk 10:10 AM – Steerable & Interactive Robots in the Era of Large Models • @DorsaSadigh 10:40 AM – ☕ Coffee Break & Poster Session 11:30 AM – All Generative Models Are Wrong, But Some Are Useful • @robo_kween 12:00 PM – Panel 1: Are LLMs, VLMs, VLAs Robust Enough for Simulation & Interaction? Yonatan Bisk, Nadia Figueroa, Roozbeh Mottaghi, moderated by @CPDArobotics 12:30 PM – Lunch 1:30 PM – Stakeholder-Informed Physical Assistance: Leveraging Generative AI towards Real-World Caregiving Robots • @TapoBhat 2:00 PM – Individualized Models of Human Control in Assistive Robotics • @brenna_argall 2:30 PM – Spotlight Talks 3:00 PM – Demonstrating Arena 5.0 • @kastner_linh 3:30 PM – ☕ Coffee Break 4:10 PM – Panel 2: Foundation Models for Shared Autonomy & Assistive Robots. @V_Vanhoucke , Brenna Argall, Tapomayukh Bhattacharjee, moderated by @haroldsoh 5:00 PM – Best Paper Award & Wrap-Up

Humanoids are improving fast! With better, more usable hardware, controllers, and advances in RL and sim2real, they’re truly leveling up #ICRA2025

Looking forward to robots waltzing to Strauss beneath Vienna’s gilded domes next year! 🎻🤖 #ICRA2026

The #ICRA2025 DEBATE “Data will Solve Robotics and Automation: True or False?” was especially fun and insightful—watch and listen carefully! Aude Billard, @animesh_garg, Leslie Kaelbling, Frank Park, Daniela Rus, @RussTedrake moderated by @Ken_Goldberg

What an incredible time at #ICRA2025! 🤖 Check out these video highlights to catch all the coolest robots in action—watch and relive the fun!

Come to the Diffusion for Manipulation Session! Room 410, Thu 15:15-15:50 #ICRA2025

Robot social navigation in the wild, as seen in the airport on my way to our Social Navigation Workshop @ #ICRA2025 Attend on Friday, Room 313, schedule here: socialnav2025.pages.dev/schedule

Read this not that long ago, great read for context nytimes.com/2024/09/21/tec…

I love keynotes – every second is a meticulously crafted blend of cutting-edge technology and the art of presenting it, with everything computed and simulated on GPUs. Hear about the new models and the future of AI:

New research could enable users to correct a robot’s behavior with simple interactions, like pointing to an item or nudging a robot’s arm. This intuitive approach would allow a layperson to guide a robot right out of the box, no machine-learning required. mitsha.re/zTB150VcHAZ

In this example, a user provides a 3D sketch (green) as input. Unconditional rollouts first miss the left block, whereas policy steering helps the model explore its learned manifold to generate a successful trajectory towards the user’s goal.

The ChatGPT of robotics needs both a model (GPT) AND an interaction method (“chat”) to let users steer robot policies online, post-training. The multimodality in diffusion models is great for capturing diverse behaviors, but users need control at inference time! @felixwyw & team’s work at #ICRA2025 presents a method to steer pre-trained policies during generation. Check out yanweiw.github.io/itps/ for all the results!