Embodied AI 101

# 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 reposition itself or move objects just to see what i...

# GR-RL: Turning a Generalist VLA Policy into a Dexterous Dexterous Manipulator

Modern robotics research is increasingly centered on large, general-purpose **Vision-Language-Action (VLA)** models – foundation policies trained or pre-trained over many tasks that can then be fine-tuned or prompt...

**ReWiND represents a significant contribution to language-conditioned reward learning, achieving 2x performance improvements over baselines through novel video rewinding augmentation and progress prediction.** However, the competitive landscape has intensified rapidly: GVL achieves zero-shot rew...

# Stage-Aware Reward Modeling (SARM) for Long-Horizon Manipulation

**Summary of the source** – Chen *et al.* (Sept. 2025) introduce *SARM*, a **stage-aware reward modeling** framework for long-horizon robot tasks (e.g. cloth manipulation). The authors note that sprawling robot tasks (integra...

# RL-100: Performant Robotic Manipulation with Real-World Reinforcement Learning

**Main claim:** RL-100 introduces a three-stage learning pipeline that combines human demonstrations and reinforcement learning to achieve *unprecedented reliability* and *efficiency* in robot manipulation. The aut...

# ResMimic: Two-Stage Residual Learning for Humanoid Loco-Manipulation

ResMimic addresses the challenging problem of **humanoid whole-body loco-manipulation** – tasks like carrying, lifting or pushing objects while walking or balancing – by decoupling general motion and task-specific control. T...

# HO-Cap Dataset and Capture System

**HO-Cap** (Hand-Object Capture) is a newly released dataset and capture setup for hand-object interaction, introduced by Wang _et al._ (UT Dallas and NVIDIA) in 2024 ([arxiv.org](https://arxiv.org/abs/2406.06843#:~:text=We%20introduce%20a%20data%20capture,pla...

The path from simulated robot to real-world deployment has long been paved with domain randomization and crossed fingers. VIRAL, a new paper from NVIDIA and collaborators, achieves **54 consecutive loco-manipulation cycles** on a Unitree G1 humanoid—walking between tables, picking up objects, placing them on trays, and repeating—with zero real-world fine-tuning. What makes this work notable isn't a novel algorithm but rather a carefully documented set of training innovations that reveal why s...

**Why ENACT matters to you:** Traditional AI benchmarks often use static images or short navigation tasks, but home robots live in a **dynamic, interactive world**. ENACT is built on the idea that “intelligence arises from sensorimotor interaction, rather than passive observation”. In practice, that means your robot should not just recognize objects in a frame, but _understand_ how its own actions change the scene. ENACT tests exactly this: it asks a model to reason about sequences of robot a...

**Introduction.** Foundation models for robot manipulation have begun leveraging **egocentric human demonstration data** as a rich resource for learning complex skills. However, human video data varies widely in context and format, making it challenging to directly use at scale. The paper _“In-N-On: Scaling Egocentric Manipulation with in-the-wild and on-task Data”_ tackles this by **strategically splitting human egocentric data into two categories** – **in-the-wild** (unconstrained human act...