The robotics crowd is back at Mission Robotics.

The Embodied AI track is a monthly research forum co-hosted with Mission Robotics, in their 8,500 sq ft brick-and-timber robotics facility in the heart of the Mission — a floor shared with more than a dozen robots.

This session is built around one question: what does a robot need before it acts — richer senses or better reasoning? Reasoning-enabled VLAs now think out loud before they move. Talk 1 asks whether that thinking can be monitored and steered at runtime, and shows the uncomfortable answer: fixing the reasoning made a driving policy markedly safer and did almost nothing for manipulation. Talk 2 comes from the touch and dexterity side — what contact-rich sensing and control buy when reasoning doesn’t. Expect the room to push on where reasoning earns its cost and where fast, contact-rich control wins.

The Frontier Research Club is a curated forum for rigorous, technical discussion at the frontier of AI. We convene researchers from the frontier labs, Stanford, Berkeley, and the teams building in production to examine concrete work — papers, methods, and results — with a bias toward assumptions, evaluation methodology, failure modes, and what would count as convincing evidence.

Presentations are intentionally brief so the majority of time is reserved for questions and critique. Materials are shared in advance so the conversation starts at depth.

Agenda

5:30pm: Doors open5:30pm – 6:30pm: Networking + light dinner6:30pm – 8:00pm: Research presentations + discussion8:00pm – 8:30pm: Networking

Presenters & topics

Talk 1 — Joseph Lee, Stanford (Somil Bansal’s lab)When Reasoning Helps Action: monitoring and steering chain-of-thought in VLA policies

Can a robot policy’s chain of thought be used as a runtime safety interface? This work introduces TRUST, a value model that predicts from a partial reasoning trace whether the reasoning will end up correct, and uses that signal to monitor and selectively steer generation in frozen VLAs. On the Alpamayo 1.5 driving VLA it monitors correctness at 88.9% accuracy, lifts reasoning correctness from 75.9% to 90.0%, and cuts collision rate by 30.4% on a hard AlpaSim subset. On a manipulation VLA the behavioral effect largely disappears — which is the point: gains in reasoning correctness do not automatically become gains in action, so the paper argues policies must be evaluated on both correctability and actionability.

Pre-read: When Reasoning Helps Action: Monitoring and Steering Chain-of-Thought in Vision-Language-Action Policies — Karnik, Lee, Gupta & Bansal, Stanford (arXiv, Sept 2026). (Read the abstract and Section 1; the result to argue about is that steering a VLA’s reasoning cut collisions 30% in driving and barely moved manipulation.)

Talk 2 — Shashank Tripathi, Postdoctoral Researcher, Max Planck Institute for Intelligent SystemsWhere humans touch the world: contact as the signal for robot learning from demonstration

Before a robot can learn from a human demonstration, it has to know where the human actually made contact — and most of the video we’d like to learn from doesn’t say. Shashank’s work recovers dense 3D human-scene and human-object contact from a single image (DECO, PICO) and treats human interaction as physics rather than pixels: bodies that push, hold, lean, and support weight. That turns ordinary footage of people using their hands into the raw material for teaching policies. The feel side of tonight’s question: what contact gives a policy that vision and reasoning alone don’t, where the estimates break, and how much of it a robot with a different body can use.

Pre-read circulated to approved attendees Wednesday morning.

Want to present your work?

If you have a research paper you’d like to discuss at one of our next sessions, please submit it for consideration. ​Submit your paper here!

Who should attend

  • Robotics researchers and research engineers

  • People training or deploying policies on real hardware — manipulation, humanoids, VLA, sim-to-real, world models

  • Anyone putting reasoning inside a robot policy — chain-of-thought VLAs, runtime monitoring, steering, and the safety case around them

  • People learning robot policies from human video and demonstration — contact, hand-object interaction, retargeting, and the perception that makes it possible

  • Autonomy and robotics safety engineers who have to decide when a policy’s reasoning can be trusted

  • Founders and engineers building tooling for robot data, simulation, and evaluation

Capacity is limited.

We will take photos and short video clips for event recap and promotion. By attending, you consent to being photographed and recorded, and to the use of those images and clips by the organizers on social media and other event marketing channels.

🌐 Connect with Frontier Research Club

Hosted by

Frontier Syndicate is a venture community connecting frontier tech researchers, builders, and investors through curated convenings and early-stage capital. Across the Bay Area, we host a recurring series of research forums, builder nights, and intimate investor dinners — and back exceptional companies emerging from the labs, communities, and technical networks we convene.​Mission Robotics is 8,500 square feet in the heart of the Mission — a converted brick-and-timber factory, hardwired for hardware, where founders at the frontier of embodied AI build alongside more than a dozen robots. Home to humanoids, robot arms, a shared data-collection rig, and a frontier world model on tap, it's purpose-built for the people teaching machines to act in the physical world.