Embodied AI 101

New framework for scalable robotic reward modeling using trajectory comparisons to train general-purpose reward models.

A single generalist VLA built on Qwen3.5-4B + 1.15B DiT flow-matching action decoder that unifies manipulation, navigation, and trajectory prediction across 11 embodiments via text-described embodiment prompts. Trained in four stages and outperforms task-specific specialists on real ALOHA and sim benchmarks without per-task fine-tuning.

Extends the EXPO method with real-world RL post-training for VLAs using image observations, action chunking, DAgger, and on-the-fly Q-value maximization. Achieves 30/30 success on 8 challenging manipulation tasks with only ~19 min of RL data on average.

RoboMeter trains dense reward models from both successful and failed robot trajectories, solving a key gap in prior methods that only learn from expert demos.

Browser-hosted mobile environment with JSON state, deterministic judges, and 256 parallel rollouts. Reports +40.7 real-device points after GRPO training on 416 tasks for GUI agent development.

Introduces a method to transfer a Unitree G1 foundation policy (Gear-Sonic) to LimX Oli/Luna humanoids using only 1% of the original compute/data. Achieves fast convergence and strong tracking performance for humanoid whole-body control.

Outputs physics-engine-compatible triangle meshes directly from sparse, unposed images without Gaussian splatting or post-processing.

Releases a benchmark suite for systematically evaluating memory in Vision-Language-Action policies on tabletop manipulation tasks.

Introduces large-scale 3D world models pretrained on diverse real-world video to enable robust robotic manipulation policies that generalize beyond simulation.

New LeRobot-based bimanual pegboard manipulation dataset with 52 episodes, 30k frames, 3 camera views, and 14-DOF arms for VLA evaluation. Provides standardized benchmark for vision-language-action model assessment.

A benchmark designed to test AI models' capabilities in making accurate 3-month future predictions.

A 30-task benchmark for evaluating long-horizon planning capabilities across 16 different AI models.

AlexNet paper that sparked the modern deep learning revolution through convolutional neural networks.

Practical insights into ML pitfalls and best practices for machine learning practitioners.

Introduces an object-centric sim-to-real policy that enables zero-shot dexterous tool use on physical robots without task-specific fine-tuning. Leverages simulation data for robust real-world transfer.

Pretrains robot policies on large-scale web video to acquire dynamics and physics understanding instead of static images or VLMs. Yields faster training, better generalization, and superior dexterous manipulation results in real-world tasks.

Introduced residual connections (ResNet) enabling training of very deep networks, still widely used in modern architectures.

Introduced the Transformer architecture based purely on attention mechanisms, becoming the foundation of nearly all modern large language models.

World foundation models for video and physics prediction with SynthID watermarking for responsible AI practices. Developed in collaboration with Google DeepMind.

Introduces a three-stage pipeline that extracts a latent action space from noisy, low-quality human motion capture, then trains a high-level RL policy in simulation to compose and execute dynamic whole-body tennis skills. Achieves volleys at human-level performance on a humanoid robot.

Closed-loop framework coupling Vision-Language Models with Video Generation Models at step-level granularity. Mitigates long-horizon drift and mid-clip errors in goal-directed video reasoning for robotic planning.

First systematic survey defining World Action Models (WAMs) as embodied foundation models that jointly predict future states and generate actions. Covers architectures, data ecosystems, and evaluation protocols.

Describes end-to-end learning of a foundation model for adaptive whole-body humanoid control via massive simulation variation. Combines proprioceptive perception and policy adaptation across embodiments.

Releases an open-source MuJoCo-based benchmark with 11 dexterous tasks, low-cost teleoperation hardware, and 1.1K human demonstrations. Designed to evaluate and train modern VLA/robotic policies.

Skill library, demonstrations, and dataset for multi-modal robotic skill learning and manipulation tasks.