Embodied AI 101
by Shaoqing Tan
Latest episode
Praxis-1: Can Video Pretraining Power Real-Robot Manipulation?
Praxis-1: Can Video Pretraining Power Real-Robot Manipulation?
About this podcast
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.
Episodes
The newest 50 of 446 episodes
- Praxis-1: Can Video Pretraining Power Real-Robot Manipulation? · 13 minRunway released Praxis-1 under Apache 2.0, an open-weights world action model that leverages video pretraining to enable real-robot control policies. Bridging large-scale video…
- Gemini Robotics 2.5: Does Better Dexterity Mean Safer Robot Control? · 13 minGemini Robotics 2.5 is a Vision-Language-Action model from Google DeepMind focused on dexterous manipulation with an emphasis on real-world safety. As a VLA from a leading lab…
- RoboCurate: Does Video-Action Verification Improve VLA Training? · 12 minRoboCurate introduces a method for verifying video-action consistency in robot training data through simulator rollouts, filtering low-quality demonstrations to improve policy…
- Can NTT's Physics-Aware World Model Prevent Manipulation Failures? · 13 minNTT unveiled a world model that integrates physics equations with video generation to predict load collapses before they occur, improving shelf-stacking success rates from 73.1%…
- PEAOM: Better Sim-to-Real Table Tennis by Modeling Camera Errors · 14 minPerception-Error-Aware Observation Modeling for Zero-Shot Sim-to-Real Control of Robotic Table Tennis
- Tactile VLA Avocado Sorting: Do Continuous Features Beat Labels? · 13 minVision-language-action (VLA) policies typically have no direct channel for non-visual physical properties, such as tactile ripeness: such evidence is either dropped entirely or…
- NEO-VLA: Why Visual Readout Depth Matters for Robot Manipulation · 13 minA 2B open VLA that feeds pixels directly through a single convolutional layer into a native vision-language backbone, eliminating a separate image encoder, trained on 2.23M…
- SLA: Can Spatial Pretraining Improve VLA Robot Manipulation? · 14 minA novel 3D-aware VLA pre-training approach that aligns visual embeddings with 3D perception, trained on 1.7M episodes and 14k+ hours of robot data, achieving top performance on…
- SLA: Does 3D Pretraining Improve pi0.5 Robot Manipulation? · 13 minHKUST researchers augmented Physical Intelligence's π0.5 VLA model with 3D visual representations during training, improving spatial reasoning without added inference cost and…
- Noetic: Can Collision-Aware World Models Make Robot Navigation Safer? · 13 minNoetic demonstrated a collision-aware latent dynamics model capable of imagining 3+ steps ahead and using a safety program to avoid walls and hazards during world model rollouts…
- Pi0.5 + SLAM: What 97% Success Means for Clothing Manipulation · 13 minA fully autonomous wheeled robot system combining the PI0.5 vision-language-action model with SLAM and collision detection achieves 97% success rate picking up and putting away…
- Delta-0: Why Humanoid Foundation Models Need Correctable Control · 13 minA generalist policy for a 69-DoF humanoid robot trained on whole-body human motion data, enabling long-horizon loco-manipulation at near-human speed via a delta action space that…
- PARMA: Can Pose-Aware Corrections Improve VLA Manipulation? · 14 minPARMA: A Pose-Aware Risk Mitigation Approach for End-Effector Execution in Vision-Language-Action Models
- LiMoDE: Can Dynamic Experts Improve Lifelong Robot Manipulation? · 14 minLiMoDE: Rethinking Lifelong Robot Manipulation From a Mixture-of-Dynamic-Experts Perspective
- A-0: Scaling Robot Manipulation with 16,000 Hours of Handheld Data · 14 minA-0 is a complete robot learning pipeline (data processing, pre-training, post-training, inference) trained purely on 16k hours of human demonstration data with zero…
- VFlash: Can Parallel VLM Planning Deliver Real-Time Robot Control? · 13 minVFlash is a fast, non-autoregressive vision-language model that outputs fixed-length geometric predictions (3D trajectories, bounding boxes) in a single forward pass, achieving…
- SADP: Can Subgoal-Aware Diffusion Improve Long-Horizon Manipulation? · 14 minSADP: Subgoal-Aware Diffusion Policy for Long-Horizon Manipulation Learned From Foundation Model Generated Demonstrations
- VAMOS: Can One Navigation VLA Work for Both Wheels and Legs? · 13 minVAMOS: A Hierarchical Vision-Language-Action Model for Capability-Modulated and Steerable Navigation
- MJev-Spatial: Faster VLM Spatial Reasoning Without Generating Text · 14 minReframes VLM spatial reasoning as a classification problem using native attention matching, requiring no new architectural heads, achieving up to 5.6× faster inference with…
- VLM Latent Shaping: Can 3D Topology Improve Spatial Reasoning? · 13 minNeurIPS 2026 Spotlight paper that identifies and manipulates how VLMs internally encode 3D scene topology, improving performance on relative-direction and spatial reasoning…
- ThinkWorld: Can Human Videos Teach VLA Manipulation Policies? · 13 minThinkWorld is a VLA framework that uses Chain-of-Dynamics reasoning to learn robot policies from action-free human videos by imagining future states and verifying them against…
- Squint: Visual RL for Sim-to-Real Manipulation in 15 Minutes · 13 minSquint: Fast Visual Reinforcement Learning for Sim-to-Real Robotics
- BudSplat: Can Budgeted 3D Gaussian Splatting Preserve Scene Detail? · 13 minIntroduces a feed-forward 3D Gaussian Splatting method that operates under explicit rendering budget constraints, enabling efficient real-time 3D scene reconstruction without…
- VoLo: Why VLA Tool Orchestration Helps Long-Horizon Manipulation · 13 minVoLo wraps VLA policies as asynchronous tools called by a language model agent that monitors execution context and triggers recovery, achieving 3–5× higher success rates on the…
- I3L: Why Corrections Beat More Demos in Mobile Manipulation · 13 minI3L introduces an iterative interactive imitation learning framework specifically targeting the last-mile performance gap in long-horizon whole-body mobile manipulation policies…
- World Labs Atlas and Marble 2: Do 3DGS Portals Advance Robotics? · 12 minA real-time portal effect that seamlessly transitions from a physical hallway capture into a different 3DGS scene, demonstrating precise alignment of virtual Gaussian splat…
- DR-WM-1.0: Can Physical Rehearsal Improve Outdoor Robot Navigation? · 14 minDR-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…
- NeoWorld-3: Can Executable 3D Programs Improve Sim-to-Real Robotics? · 13 minNeoWorld-3 converts visual observations into executable 3D programs with geometry refinement and physics validation, enabling real-to-sim-to-real robot transfer tested across…
- Genie 4 World Simulation: Is It Ready for Robot Training? · 12 minDeepMind 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…
- SuperMimic: Training Humanoid Motion Control on One GPU · 13 minSuperMimic 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…
- RynnUnify: Does Video Imagination Improve Robot Planning? · 13 minA 16-billion parameter unified embodied model that integrates long-horizon visual planning, physical reasoning, and robot control through interleaved video-language reasoning…
- LaLa-World: Can Language Scale Latent World Models for Robots? · 14 minA 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…
- SwarmACB: Why Behavior Selection Helps Swarm Sim-to-Real Transfer · 13 minReinforcement learning offers a powerful framework for automatically generating control policies for robot swarms, yet its application to physical multirobot systems has been…
- GVC: Why Action Modeling Helps Robot Reinforcement Learning · 13 minGVC parametrizes Q-functions as generative value classifiers that avoid action conditioning, enabling better integration with large pre-trained robot policies and richer learning…
- World Labs Atlas: Is Single-Image 3D Generation Ready for Robotics? · 13 minWorld 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…
- SpatialDex: Does 3D Video Pretraining Help Bimanual Manipulation? · 13 minSpatialDex 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…
- Show-Harness: Can Small VLMs Replace Frontier Models in Robot Control? · 13 minShow-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…
- MolmoBot: Can Simulation Alone Deliver Zero-Shot Manipulation? · 13 minMolmoBot leverages large-scale simulation to enable zero-shot manipulation transfer to real robots without task-specific training data. This advances the sim-to-real…
- VLA-Δ: Can World-Change Supervision Improve Robot Manipulation? · 13 minVLA-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…
- Axis Robotics Browser Simulation: Does Sim-to-Real Manipulation Work? · 13 minA 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…
- HexaShell: Can Debugging Code Generalize RGB Robot Manipulation? · 12 minHexaShell 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…
- Niantic Spatial Gaussian Splats: Better Sim-to-Real Robot Training? · 12 minNiantic Spatial released high-fidelity 3D Gaussian splat scenes with aligned meshes and collider-ready geometry for warehouse robot simulation and policy training. By combining…
- Gemini 4 Argon for Robotics: What the Benchmarks Actually Show · 12 minGoogle released Gemini 4 Argon, a physical reasoning model that scores 87% on the ERQA embodied-reasoning benchmark using pure reasoning, outperforming other frontier models…
- GreenWAM: Can One World-Action Model Learn Across Robot Bodies? · 13 minGreenWAM 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…
- 3DCodeVerse: Can Code Generation Build Better Worlds for Robotics? · 14 min3DCodeVerse introduces a model trained on 2M programs that generates static and articulated 3D objects, scenes, and entire worlds represented as executable code. The…
- Motion- and Gravity-Aware 3D Reconstruction: Do Offline Priors Help? · 13 minA 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…
- ASPIRE: Can Code Refinement Make Robot Manipulation Generalize? · 13 minNVIDIA's ASPIRE framework combines GPT-6 Astra with continual skill refinement to achieve saturated performance (99–100%) on LIBERO, Robosuite, and BEHAVIOR-1K simulation…
- UAV-ESPrune: Can Visual Token Pruning Preserve UAV Reasoning? · 13 minLow-altitude unmanned aerial vehicles (UAVs) require compact multimodal inference for perception, environment understanding, and planning-oriented reasoning. Vision-language…
- Gemini Robotics 2: How General Is Its Whole-Body Robot Control? · 30 minGoogle DeepMind's Gemini Robotics 2 demonstrates whole-body control, fine-grained dexterity, and embodied reasoning across multi-step real-world manipulation tasks. It highlights…
- RoboJEPA to Long-WAM: Do New Embodied AI Models Generalize? · 14 minA cluster of new releases including RoboJEPA (latent world models spanning 12 robot embodiments), UNITAS (3D-native world-action model), and DreamTrue (failure prediction via…
