MIT CSAIL

Robots lack rich & diverse simulation content, which could teach them tasks across various environments. "SceneSmith" uses three AI agents to create more realistic 3D scenes of places like kitchens & hotels, saving roboticists time on real-world testing: bit.ly/4wkaf5r

Robots lack rich & diverse simulation content, which could teach them tasks across various environments. "SceneSmith" uses three AI agents to create more realistic 3D scenes of places like kitchens & hotels, saving roboticists time on real-world testing: bit.ly/4wkaf5r

Excited to share that our work NeuralActuator: Neural Actuation Modeling for Robot Dynamics and External Force Perception has been accepted to #RSS2026! Your robot — even a low-cost one — can feel external forces without torque or tactile sensors. TL;DR: NeuralActuator is a neural actuator model that jointly predicts 1️⃣torque to capture the nonlinear and time-varying current–to–torque relationship of low-cost servos, 2️⃣external contact forces (and force detection gates) for sensorless force perception, 3️⃣and motor conditions that indicate each motor’s operating regime. Here is a fast-forward video clip ⬇️ We are also covering more robots like LeRobot-S101 and Franka Panda. More details coming soon.

"The first step of any project is to grossly underestimate its complexity and difficulty." — Nicoll Hunt

A free guide that answers key reinforcement learning questions: shorturl.at/zcciv v/@arjunkocher & @sheriyuo

Happy birthday, Nikola Tesla! He helped pioneer the alternating current (AC) power systems that became the foundation of modern electric power distribution. Image v/Culver Pictures

In both projects, researchers are helping ML systems learn from more realistic 3D environments to better simulate real-world scenarios. The core idea is to capture the physical world virtually, making 3D testing more efficient & reducing the need for real-life trial & error.

"GeoPT" 🛬 GeoPT teaches simulation models physics by reenacting mechanical interactions in 3D, reaching peak performance 2x faster w/up to 60% less data. It could help predict how vehicles, objects, & robots handle wind, water, & collisions: bit.ly/4voRGM6

MIT CSAIL is at ICML this week. One common thread: Teaching ML systems in realistic virtual environments🧵 A couple of examples ⬇️

MIT’s self-organizing boats can create flexible infrastructure on demand. The team's "FloatForm" system uses a swarm of small aquatic robots, turning water into a programmable surface. These temporary platforms could help w/urban planning & emergencies: bit.ly/4vk61JA

If you're attending ICML this week & are interested in learning more about either project, send us a DM. You can find more info about CSAIL research on our website: bit.ly/4dQ6uOA

Some more recommended reads, each w/MIT CSAIL authors.

How can robots learn to do tasks when instructions are ambiguous? MIT’s "Masked IRL" uses LLMs to help robots understand vague chore requests & focus on the key details. One model clarifies users’ prompts, then another ignores irrelevant info: bit.ly/4vQnzhH

Free tutorials to help you learn programming tools, v/@ComputerrBooks.

What's your favorite generative AI tool?

What came before LaTex, v/@Anthony_Bonato.

Massive synaptic pruning between birth and adulthood: the nervous system gets simpler with age! Two-decade connectomic effort in Lichtman lab @harvard reveals how the order in which motor neurons activate decides which connections survive. @NatureNeuro 👉 nature.com/articles/s4159…

Ask us your questions about human-robot collaboration. We’ll pick some to feature in an upcoming 32 questions video w/MIT PhD student Alex Cuellar.