# Robo-Dopamine: Dense Reward Learning for High-Precision Robot Manipulation In a recent preprint (Tan et al., *Robo-Dopamine: General Process Reward Modeling for High-Precision Robotic Manipulation*), researchers from Peking University, BAAI, Unive...

# AstraNav-Memory: Compressing Visual Context for Lifelong Navigation Recent work by Botao Ren, Junjun Hu, Xinda Xue and others (Alibaba’s Amap, Tsinghua, Peking) introduces **AstraNav-Memory: Contexts Compression for Long Memory**, a method for giv...

# DexWM: World Models for Dexterous Manipulation from Human Videos Dexterous manipulation – the art of using multi-fingered hands to pick, place, twist, or otherwise manipulate objects – remains a towering challenge in robotics. Even seemingly simp...

# Video-Action Models: Generalizing Robot Control with Video Diffusion In a new preprint titled *“mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs”* (Pai et al., arXiv 25 Dec 2025), Jonas Pai and colleagues (from Mimic Ro...

# Calibrated Confidence in Controllable Video Models In the recent preprint *“World Models That Know When They Don't Know: Controllable Video Generation with Calibrated Uncertainty”* (Mei et al., 2025) ([huggingface.co](https://huggingface.co/papers...

# Scaling Up Offline Model-Based RL with Action Chunks (MAC) In *“Scalable Offline Model-Based RL with Action Chunks”* (Park et al., 2025) – by Kwanyoung Park, Seohong Park, Youngwoon Lee, and Sergey Levine (UC Berkeley and Yonsei Univ.) – the autho...

# Emergent Human-to-Robot Transfer in Vision-Language-Action Models **Simar Kareer, Karl Pertsch, James Darpinian, Judy Hoffman, Danfei Xu, Sergey Levine, Chelsea Finn, Suraj Nair Physical Intelligence (PI); Georgia Tech – *“Emergence of Human to ...

# DexScrew: Learning Dexterous Manipulation from “Imperfect” Simulations *Learning Dexterous Manipulation Skills from Imperfect Simulations* is a recent preprint by Hsieh *et al.* (Dec 2025) from UC Berkeley (Elvis Hsieh*, Wen-Han Hsieh*, Yen-Jen Wa...

# X-Humanoid: Robotizing Human Videos into Humanoid Videos A new preprint titled **“X-Humanoid: Robotize Human Videos to Generate Humanoid Videos at Scale”** (Pei Yang, Hai Ci, Yiren Song, Mike Zheng Shou, NUS, Dec 2025) tackles a fundamental data b...

# Decoupled Q-Chunking: Combining Long-Range Value Propagation with Reactive Policies In *Decoupled Q-Chunking* (Li, Park, Levine, 2025), the authors tackle a long-standing issue in reinforcement learning: how to propagate value efficiently over far...

# F2D2: Joint Distillation for Fast Likelihood and Sampling in Flow Models In the recent preprint **“Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models”** by Xinyue Ai, Yutong He, Albert Gu, Ruslan Salakhutdinov, J...

# Training-Time Action Conditioning for Efficient Real-Time Chunking In “Training-Time Action Conditioning for Efficient Real-Time Chunking” (Black et al., 2025), researchers from Physical Intelligence (including Kevin Black, Allen Z. Ren, Michael E...

In a recent technical report, the Physical Intelligence team (led by Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Kevin Black, Ken Conley, Grace Connors, and dozens of others) introduces **“π0.6: a VLA That Learns From Experience.”** The key ide...

# Robotic World Model (RWM): Learning Stable Neural Simulators for Long-Horizon Control Recent work by Chenhao Li, Andreas Krause, and Marco Hutter (ETH Zurich) introduces **Robotic World Model (RWM)** – a learned, recurrent simulator designed to su...

# Much Ado About Noising: Dispelling the Myths of Generative Robotic Control In recent years, robotics and control researchers have widely embraced *generative control policies* – that is, parameterizations of robotic policies using flow or diffusio...

# ReWiND: Language-Guided Reward Learning for Robot Task Adaptation Modern robot learning systems crave rich supervision, but collecting demonstrations or hand-designing rewards for every new task is impractical. Imagine wanting a robot to stack dif...

# Nested Learning: Unifying Architecture and Optimization for Memoryful Models In “Nested Learning: The Illusion of Deep Learning Architectures” (Behrouz *et al.*, NeurIPS 2025), researchers at Google Research (with Peilin Zhong at Columbia) propose...

# AsyncVLA: Real-Time Vision-Language-Action through Asynchronous Flow Matching **Context & Motivation.** Robotics is riding a wave of foundation models and vision-language-action (VLA) controllers that promise generalist “pick-and-place” skills fo...

# GigaWorld-0: World Models as a Data Engine for Embodied AI **1. What Problem Is This Paper Trying to Solve?** Modern embodied AI — controlling robots and agents via vision and language — is bottlenecked by data. Training vision‐language‐action (V...

# AAWR: Using Privileged Training to Learn Active Perception in Robotics Modern robots often must act under partial observability: their onboard sensors don’t immediately reveal all task-relevant information. For instance, a robot may have to repos...