Say ahoy to 𝚂𝙰𝙸𝙻𝙾𝚁⛵: a new paradigm of *learning to search* from demonstrations, enabling test-time reasoning about how to recover from mistakes w/o any additional human feedback! 𝚂𝙰𝙸𝙻𝙾𝚁 ⛵ out-performs Diffusion Policies trained via behavioral cloning on 5-10x data!

⛵️ Excited to share 𝚂𝙰𝙸𝙻𝙾𝚁: a method for *learning to search* with learned world + reward models to plan in the latent space at test-time. Unlike behavior cloning, 𝚂𝙰𝙸𝙻𝙾𝚁 recovers from mistakes without any additional data, DAgger corrections, or ground truth rewards.

Behavior cloning hits a wall even with more data BUT RL can break through this wall! Simply scaling up teleoperation data for robotics seems unlikely to work. A big fan of methods such as SAILOR that can scale up search instead of relying on more expert data 🚀

Interesting work on learning more efficiently from demos. Learns a world models that allows search at inference time.

My school friend, Kaushik’s son ‘Vibhakar Mohta’ is an integral part of this team. These guys are doing some phenomenal path breaking work in the field of AI. Vibhakar Mohta ✈️ NeurIPS 2025

All hands on deck for SAILOR ⚓️! 🎣 SAILOR nets more performance even with 10x less data than Diffusion Policies trained with behavioral cloning. Check out Gokul Swamy’s post explaining how we leverage learned models to enable test-time planning and mistake recovery for robots!

Really cool work from Gokul Swamy ✈️ NeurIPS 2025 and co! Sample efficiency gains by leveraging search in world models is a cool way to enable test time reasoning for robotics!

"Learning to search" as a way to improve scaling from robot data

Finally get to share one of the coolest codebases we built this past year! We combine search with learnt reward and world models to significantly outperform diffusion policies with very little demonstration data. Check out Gokul Swamy super deep post and download our code today!

The entire team grinded on this, but I have been blown away by the rigor that Arnav Jain ✈️ Neurips 2025 and Vibhakar Mohta put into running experiments. And of course Gokul Swamy ✈️ NeurIPS 2025 for masterminding this as always. Very optimistic that this, alongside our sim2real efforts is the final playbook.

CMU and Cornell researchers released SAILOR which allows for test-time reasoning and beats other diffusion policies trained with 5-10x more data! Congrats Gokul Swamy ✈️ NeurIPS 2025 and team x.com/g_k_swamy/stat…

Teleoperation is slow, expensive, and difficult to scale. So how can we train our robots instead? Introducing X-Sim: a real-to-sim-to-real framework that trains image-based policies 1) learned entirely in simulation 2) using rewards from human videos. portal-cornell.github.io/X-Sim

.Lucid Motors’ premium EVs. @Nuro’s proven L4 autonomy. Uber’s global ride-hailing network. Together, we're launching a next-gen robotaxi fleet—20K+ vehicles, starting in 2026. Details here: nuro.ai/nuro-lucid-ube… #autonomousvehicles #technology #innovation #partnership

Stay tuned!

Full episode dropping soon! Geeking out with Gokul Swamy Arnav Jain Vibhakar Mohta on 𝚂𝙰𝙸𝙻𝙾𝚁: Robust Imitation via Learning to Search gokul.dev/sailor/ Co-hosted by Michael Cho - Rbt/Acc Chris Paxton

Full episode dropping soon! Geeking out with Gokul Swamy Arnav Jain Vibhakar Mohta on 𝚂𝙰𝙸𝙻𝙾𝚁: Robust Imitation via Learning to Search gokul.dev/sailor/ Co-hosted by Michael Cho - Rbt/Acc Chris Paxton

Thank you soo much for having us!

Excited to be featured in robopapers!

Best Self-Improving Agent: ReviveAgent Shaurya Rohatgi @ NeurIPS 2025 Varad Pimpalkhute Vibhakar Mohta They built an agent that acts as a dedicated maintenance engineer for your codebase. Its custom 'self-reflection' loop allows it to learn from its own failures, getting smarter at resolving and refactoring

⛵️ Setting sail at #NeurIPS2025! Excited to present SAILOR as a Spotlight poster this week. 🗓 Thursday @ 11 AM 🔗 gokul.dev/sailor/ If you're around, let's grab a coffee ☕ and discuss about world Models, flow matching, diffusion, VLAs & RL !