Skip to content
Sign in
Artwork for Embodied AI 101
TechnologyScience

Embodied AI 101

Shaoqing Tan

Embodied AI 101 tracks the latest research at the intersection of AI and physical intelligence.

Robotics, manipulation, world models, and the path from digital intelligence to embodied agents — each episode breaks down recent papers and advances in the field.

We examine the methodology, real-world implications, and path to deployment. Covering robot learning, perception, control, and the fundamental challenge of grounding intelligence in the physical world.

New episodes daily.

Play
Open in
  • 118 episodes
  • daily
  • Avg 25 min
  • English

Episodes 118

NewestOldestLongest
All20Unplayed20In progress0Played0
Counted on this page — what you have heard stays on this device, so it is not something the list can be paged by.
  • October 17 · 12 min

    Motion- and Gravity-Aware 3D Reconstruction: Do Offline Priors Help?

    A new feed-forward 3D reconstruction method injects optical flow and gravity priors directly into the network to robustly handle dynamic scenes and motion blur without any additional inference cost. This advances camera-based spatial intelligence for embo

    • Transcript
  • October 17 · 13 min

    ASPIRE: Can Code Refinement Make Robot Manipulation Generalize?

    NVIDIA's ASPIRE framework combines GPT-6 Astra with continual skill refinement to achieve saturated performance (99–100%) on LIBERO, Robosuite, and BEHAVIOR-1K simulation benchmarks, with open-sourced code targeting real-world manipulation transfer. The r

    • Transcript
  • October 16 · 13 min

    UAV-ESPrune: Can Visual Token Pruning Preserve UAV Reasoning?

    Low-altitude unmanned aerial vehicles (UAVs) require compact multimodal inference for perception, environment understanding, and planning-oriented reasoning. Vision-language models (VLMs) provide a shared interface for these tasks, but long visual-token s

    • Transcript
  • October 15 · 30 min

    Gemini Robotics 2: How General Is Its Whole-Body Robot Control?

    Google DeepMind's Gemini Robotics 2 demonstrates whole-body control, fine-grained dexterity, and embodied reasoning across multi-step real-world manipulation tasks. It highlights the growing importance of rich spatial and environmental data for scaling ro

    • Transcript
  • October 15 · 13 min

    RoboJEPA to Long-WAM: Do New Embodied AI Models Generalize?

    A cluster of new releases including RoboJEPA (latent world models spanning 12 robot embodiments), UNITAS (3D-native world-action model), and DreamTrue (failure prediction via robot world models) collectively push forward the frontier of embodied world mod

    • Transcript
  • October 15 · 29 min

    VLA-ACL: Can Token Pruning Speed Up Robot Manipulation?

    VLA-ACL (Action Consistency Learning) learns a lightweight visual token pruning policy through action-level supervision while keeping the base VLA model entirely frozen. The training objective encourages actions produced from pruned visual contexts to rem

    • Transcript
  • October 15 · 28 min

    Juno: Do Predictive Latents Make VLA Manipulation More Robust?

    Juno introduces a unified framework built around one action-conditioned JEPA that serves as a control-aligned representation backbone, a predictive teacher, and an adaptable dynamics model for VLA models. It addresses three failures: mismatch with embodim

    • Transcript
  • October 15 · 27 min

    Can Rephrasing Fix VLA Language Sensitivity Without Retraining?

    Vision-language-action models (VLAs) are strikingly sensitive to instruction phrasing and do not inherit the language robustness of the vision-language models they are built on. A one-word edit can move success by tens of points: pi0.5 turns on a LIBERO s

    • Transcript
  • October 14 · 28 min

    Can CriticGPT Replace Hand-Designed Rewards for Robot Manipulation?

    Learning scalable reward signals for robotic manipulation remains challenging. CriticGPT uses a multimodal LLM as a critic to provide reward signals for robot manipulation, avoiding the need for manually designed dense rewards or costly human preference a

    • Transcript
  • October 14 · 28 min

    Urbex: Can Agentic VLM Search Improve City-Scale 3D Grounding?

    Urbex reframes city-scale 3D grounding as an active spatial search problem, using an agentic vision-language model that iteratively queries landmarks, zooms into regions, and renders oblique views to localize targets. The approach outperforms prior baseli

    • Transcript
  • October 12 · 28 min

    VPP2: Does Better Video Prediction Improve Robot Manipulation?

    World action models (WAMs) have emerged as an important class of generalist robot policies, aiming to transfer video prediction priors to action learning. Finds that existing WAMs frequently fail and proposes improvements.

    • Transcript
  • October 12 · 30 min

    PointWAM: Does 3D World Modeling Improve Dexterous Manipulation?

    Introduces a 3D world action model that forecasts scene and hand trajectories as point clouds, pre-trained on 1.15M human videos. Achieves state-of-the-art performance on the DexJoCo benchmark for dexterous manipulation, advancing how embodied agents pred

    • Transcript
  • October 11 · 27 min

    Mulligan: Can Failure-Focused Data Improve Robot Imitation Learning?

    Mulligan targets failure-prone initial states during real-robot deployment rather than uniform sampling, making DAgger-style imitation learning significantly more data-efficient. Validated across 3 real tasks with 2,550 blind evaluations, it improves poli

    • Transcript
  • October 11 · 28 min

    SOLE-R1: Can Video-Language Rewards Train Manipulation Policies?

    SOLE-R1 trains a video-language reasoning model to generate dense reward signals directly from real robot video, enabling online reinforcement learning on manipulation tasks from scratch without hand-crafted reward functions. This removes a key bottleneck

    • Transcript
  • October 10 · 30 min

    Odyssey-3: Does a Physics-IQ World Model Transfer to Real Robots?

    Odyssey-3 achieves a new state-of-the-art Physics-IQ score of 66.1, generating interactive, physically plausible worlds that transfer to real robots, autonomous vehicles, and drones, enabling recovery behaviors with limited real-world data. It represents

    • Transcript
  • October 10 · 28 min

    Long-WAM: Does Longer Visual Context Improve Real-Time Robot Control?

    Long-WAM is a world-action model that extends visual context length for real-time robot control, showing that longer video histories yield larger performance gains when the video foundation model is pretrained autoregressively. This advances robot foundat

    • Transcript
  • October 9 · 29 min

    WR1.0: Does Whole-Body Humanoid Control Prove Generalization?

    WR1.0 is a whole-body humanoid foundation model demonstrating coordinated locomotion combined with dexterous manipulation tasks such as folding clothes, representing a significant step toward unified motor control across a humanoid body. The demo highligh

    • Transcript
  • October 9 · 28 min

    WorldGuide: Better Video Planning, Not Proven Robot Control

    WorldGuide is a video world model designed for goal-directed procedural task execution, enabling embodied agents to plan and execute multi-step tasks by predicting future visual states conditioned on goals. It advances the use of world models as explicit

    • Transcript
  • October 8 · 29 min

    Humanoid Scooter Control: Does Sim-to-Real Reinforcement Learning Work?

    This paper presents a simulation-to-reality (sim-to-real) transfer process, allowing a full-sized humanoid robot to autonomously balance a two-wheeled scooter and track operator-supplied heading commands through a deep reinforcement learning (DRL) approac

    • Transcript
Showing 21–40 of 118 episodes