Robotics & World Models Reading Club 03: Robotic Policy Adaptation — San Francisco

​Robotics & World Models Reading Club 03: Robotic Policy Adaptation — San Francisco

​A high-signal reading group for AI researchers & builders pushing the frontiers of robotic world models, WAMs, and embodied intelligence. In our previous sessions, we brought together researchers and engineers from Boston Dynamics, Google, NVIDIA, Stanford, UC Berkeley and leading Bay Area robotics startups.

​​Hosted by Junfan Zhu & Aurora Feng.

​​​Supported by Neural Motion, a universal cross-embodiment data representation layer for embodied AI.

​Reading Club 03's Core Theme

​Robotic Policy Adaptation: Fast and Reliable Bridging of Deployment Gaps in the Foundation Model Era

​Adapting robotic policies to new environments remains one of the last unsolved problems in robotics. Real-world deployment rarely matches training conditions, making robust, fast adaptation a fundamental bottleneck.

​This Saturday, Haoyi Niu (UC Berkeley) will share his research on algorithms for efficient cross-domain policy adaptation under real-world constraints.

​Beyond technical methods, this keynote offers a candid look into research meta-cognition:

• ​Why focus on policy adaptation in the era of foundation models?

• ​How to stay "on-trend" with fast-moving AI trends while maintaining a long-term research thesis?

​Pre-Readings

• ​H2O: When to Trust Your Simulator: Dynamics-Aware Hybrid Offline-and-Online Reinforcement Learning

  • ​A foundational framework for hybrid offline–online RL under sim-to-real gaps.

  • ​H2O introduces a dynamics gap estimator ω(s, a), modeling the discrepancy between real-world and simulated state-action distributions. This estimator is integrated into a dynamics-aware policy evaluation objective, where simulated transitions are adaptively down-weighted in high-gap regions.

  • ​Key insight:
    → Prevents policies from exploiting simulator artifacts while still leveraging scalable simulation data.

• ​H2O+: An Improved Framework for Hybrid Offline-and-Online RL with Dynamics Gaps

  • ​A more modular and scalable extension of H2O.

    • ​Decouples offline and online components into plug-and-play modules

    • ​Supports arbitrary offline RL + online RL combinations

    • ​Improves stability and sample efficiency under large dynamics mismatch

  • ​Key shift:
    → From a fixed framework → a general-purpose hybrid learning system

• ​xTED: Cross-Domain Adaptation via Diffusion-Based Trajectory Editing

  • ​A data-centric alternative to policy adaptation.

  • ​Instead of adapting policies, xTED adapts trajectories:

    • ​Train a diffusion model on target-domain trajectories

    • ​Add noise to source trajectories → denoise with target prior

    • ​Preserve task semantics while injecting target-domain dynamics

  • ​Result:
    → Zero/few-shot transfer without policy retraining

• ​RL-Token: Precise Manipulation with Efficient Online RL (Physical Intelligence, 2026)

  • ​Bridging large VLA models with real-time robotic learning.

    • ​Introduces a compact readout token distilled from frozen VLA representations

    • ​Uses lightweight actor-critic RL on real robots

    • ​Edits (not replaces) VLA actions for stability

  • ​Performance:
    → Up to 3× speedup in high-precision tasks (e.g., insertion, fastening)
    → Adaptation in minutes to hours

​Location

​​​San Francisco (Downtown)

​​Date & Time

​​​Saturday, April 11, 2026 | 2:00 PM – 5:00 PM

​​​Join Discord Community

​​​https://discord.gg/WH7DrTHRXK

​​Follow Saturday Robotics on X

​​https://x.com/saturdayrobotic

​Agenda

​​2:00 PM – 2:30 PM Door Opens & Social

• ​​​Food 😋, beverages🧋 and UNLIMITED strawberries 🍓 (our official reading club fruits ☺️😄).

​​2:30 PM – 3:00 PM "Robotic Policy Adaptation" Keynote by Haoyi Niu, UC Berkeley (https://t6-thu.github.io/)

​​​Virtual Keynote via Zoom: TBD

​Recording: TBD (We are looking for recording volunteers)

​​3:00 PM – 5:00 PM Q&A, ​open-floor roundtable (10–20 min per topic) on spotlight papers or any paper you’d like to highlight. Feel free to share why the paper matters and its technical details.

​Past events

​#⁠reading-club-02-0404

• ​​Session 02 Luma: https://luma.com/g3qrrti0

• ​​Reading Club 02 Review (liked by Yann LeCun on X): https://x.com/junfanzhu98/status/2040716119259164673?s=20

• ​​Event photos: https://x.com/junfanzhu98/status/2040717084972245341?s=20

• ​​LinkedIn more photos: https://www.linkedin.com/posts/junfan-zhu_robotics-world-model-reading-club-02-hot-ugcPost-7446465600723509248-1qbj?utm_source=share&utm_medium=member_desktop&rcm=ACoAABxP-p0BpUNGDf347aKh_1uJAPzG4er0As8

​#⁠reading-club-01-0328

• ​​​Session 01 Luma: https://luma.com/8s4w1wu6

• ​​​Reading Club 01 Review: https://x.com/junfanzhu98/status/2038153945219305812

• ​​​Event photos (liked by Yann LeCun on X): https://x.com/junfanzhu98/status/2038161288090779985

​​Logistics

​Spots are limited. Please arrive by 2:00 PM for check-in. Keynote will begin promptly at 2:30 PM.

• ​​We currently do not have volunteers available to assist with late check-ins. Given the high volume of inquiries and 100+ attendees (both online and onsite), we kindly ask that you arrive on time to ensure smooth entry.