Skip to content
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
  • 159 episodes
  • Avg 22 min
  • English
Counted on this page — what you have heard stays on this device, so it is not something the list can be paged by.
  • October 31 · 13 min

    I3L: Why Corrections Beat More Demos in Mobile Manipulation

    I3L introduces an iterative interactive imitation learning framework specifically targeting the last-mile performance gap in long-horizon whole-body mobile manipulation policies. By incorporating human feedback loops during policy refinement, the method a

    • Transcript
  • October 30 · 11 min

    World Labs Atlas and Marble 2: Do 3DGS Portals Advance Robotics?

    A real-time portal effect that seamlessly transitions from a physical hallway capture into a different 3DGS scene, demonstrating precise alignment of virtual Gaussian splat content with live camera views using World Labs Atlas and Marble 2. Illustrates a

    • Transcript
  • October 29 · 13 min

    DR-WM-1.0: Can Physical Rehearsal Improve Outdoor Robot Navigation?

    DR-WM-1.0 is a world model trained on real terrain physics to enable physical rehearsal for wheel-legged robots, complemented by DR-VLA-2.0 and the EgoLynx/LynxScenes outdoor benchmark suite. This advances sim-to-real transfer and vision-language-action p

    • Transcript
  • October 29 · 13 min

    NeoWorld-3: Can Executable 3D Programs Improve Sim-to-Real Robotics?

    NeoWorld-3 converts visual observations into executable 3D programs with geometry refinement and physics validation, enabling real-to-sim-to-real robot transfer tested across 212+ objects and scenes. This advances how robots can build and exploit structur

    • Transcript
  • October 28 · 12 min

    Genie 4 World Simulation: Is It Ready for Robot Training?

    DeepMind released Genie 4, a real-time interactive world model powered by Gemini that generates steerable simulations with native audio, music, and dialogue from natural language inputs. This represents a significant leap in generative world modeling, ena

    • Transcript
  • October 28 · 12 min

    SuperMimic: Training Humanoid Motion Control on One GPU

    SuperMimic scales humanoid motion control across 129k diverse motions using a single GPU in 3 hours, compared to prior methods requiring 128 GPUs over 7 days, dramatically reducing the compute barrier for humanoid policy training. This efficiency breakthr

    • Transcript
  • October 27 · 12 min

    RynnUnify: Does Video Imagination Improve Robot Planning?

    A 16-billion parameter unified embodied model that integrates long-horizon visual planning, physical reasoning, and robot control through interleaved video-language reasoning. Represents a significant scale-up for generalist robot foundation models target

    • Transcript
  • October 27 · 13 min

    LaLa-World: Can Language Scale Latent World Models for Robots?

    A world model that encodes both actions and visual state changes as text, enabling a single latent model to train jointly on simulation, robot, and video data. This unified representation accelerates planning and policy fine-tuning across diverse data sou

    • Transcript
  • October 26 · 13 min

    SwarmACB: Why Behavior Selection Helps Swarm Sim-to-Real Transfer

    Reinforcement learning offers a powerful framework for automatically generating control policies for robot swarms, yet its application to physical multirobot systems has been limited by the simulation-to-reality gap. Despite substantial...

    • Transcript
  • October 25 · 12 min

    GVC: Why Action Modeling Helps Robot Reinforcement Learning

    GVC parametrizes Q-functions as generative value classifiers that avoid action conditioning, enabling better integration with large pre-trained robot policies and richer learning signals from pretraining data. This approach advances offline and online rei

    • Transcript
  • October 24 · 12 min

    World Labs Atlas: Is Single-Image 3D Generation Ready for Robotics?

    World Labs released beta access to Atlas, a spatial intelligence model that builds large-scale 3D environments and flythrough videos from a single photograph. This represents a significant advance in single-image 3D scene understanding and generation at s

    • Transcript
  • October 24 · 13 min

    SpatialDex: Does 3D Video Pretraining Help Bimanual Manipulation?

    SpatialDex reconstructs scenes and hands in scale-aligned 3D from 1M RGB human videos and trains bimanual manipulation policies that nearly double prior VLA success rates on real robot tasks. The approach demonstrates how 3D spatial grounding from large-s

    • Transcript
  • October 23 · 12 min

    Show-Harness: Can Small VLMs Replace Frontier Models in Robot Control?

    Show-Harness fine-tunes a lightweight VLM (e.g., Qwen3.5-2B) in hours to serve as a capable local robot agent, eliminating the need for large models at inference time. This approach makes VLM-based robot intelligence practical for on-device deployment wit

    • Transcript
  • October 22 · 13 min

    MolmoBot: Can Simulation Alone Deliver Zero-Shot Manipulation?

    MolmoBot leverages large-scale simulation to enable zero-shot manipulation transfer to real robots without task-specific training data. This advances the sim-to-real generalization frontier for vision-language-action models in manipulation.

    • Transcript
  • October 22 · 13 min

    VLA-Δ: Can World-Change Supervision Improve Robot Manipulation?

    VLA-Δ trains robot policies on how the physical world changes rather than raw actions, matching state-of-the-art on LIBERO and RoboTwin benchmarks and outperforming π0.5 on real robots. This world-change modeling paradigm offers a novel perspective on gro

    • Transcript
  • October 21 · 12 min

    Axis Robotics Browser Simulation: Does Sim-to-Real Manipulation Work?

    A browser-based teleoperation interface enables collection of robot arm demonstrations in simulation that transfer to real hardware (e.g., Franka) via sim-to-real, achieving high success rates with mixed sim and few real demos. This lowers the barrier to

    • Transcript
  • October 19 · 12 min

    HexaShell: Can Debugging Code Generalize RGB Robot Manipulation?

    HexaShell treats robot control as a coding task where an agent writes reusable tools from RGB camera observations, achieving 39.7% success across 42 tasks on the RoboDojo benchmark — a new SOTA. This matters because it demonstrates that code-as-action abs

    • Transcript
  • October 19 · 12 min

    Niantic Spatial Gaussian Splats: Better Sim-to-Real Robot Training?

    Niantic Spatial released high-fidelity 3D Gaussian splat scenes with aligned meshes and collider-ready geometry for warehouse robot simulation and policy training. By combining photorealistic appearance with physically accurate collision geometry, this ap

    • Transcript
  • October 19 · 12 min

    Gemini 4 Argon for Robotics: What the Benchmarks Actually Show

    Google released Gemini 4 Argon, a physical reasoning model that scores 87% on the ERQA embodied-reasoning benchmark using pure reasoning, outperforming other frontier models. Argon controls robots not through raw float trajectories but via a code-as-polic

    • Transcript
  • October 18 · 12 min

    GreenWAM: Can One World-Action Model Learn Across Robot Bodies?

    GreenWAM is an open-weights world-action foundation model that jointly predicts robot actions and selects them across 20 robot families within a single unified model. Its cross-embodiment generalization and open release make it a significant step toward s

    • Transcript
Showing 41–60 of 159 episodes