Embodied AI 101

Proposes a unified RL policy trained in simulation that jointly handles force and position control without force sensors, enabling compliant behaviors and improved force-aware imitation learning on quadrupedal and humanoid robots.

A single vision-language model that unifies high-level reasoning, long-horizon planning, and human-robot interaction, serving as a cognitive layer above low-level vision-language-action models.

A 3B-parameter open-source unified decoder-only transformer for general robot control that jointly handles perception, planning, reasoning, and action via interleaved vision-text-action pretraining on a 1.5M-sample dataset. Achieves strong results across manipulation tasks and benchmarks with fully open-sourced model, code, and dataset.

Introduces Video-Action Models (VAMs) that replace standard VLMs in VLAs with a pretrained internet-scale video backbone (Cosmos-Predict2) combined with a flow-matching action decoder. Achieves approximately 10× better sample efficiency on real-world pick-and-place tasks.

An open-source, sub-$500 7-DOF 3D-printed robotic arm with 1.2 kg payload, wrist and head cameras, and full LeRobot integration, designed to lower the barrier to entry for real manipulation research.

A 4D embodied world model that predicts RGB, depth, and optical flow from RGB-D input and instructions using a tri-branch diffusion architecture. Enables closed-loop bimanual robot control by bridging world prediction and policy execution.

A digital teleoperation system that uses hand-pose streams to drive a world model for real-time (40+ FPS) high-fidelity robot video rendering from a single image. Removes physical robots from the training loop entirely.

A lightweight plug-in that monitors latent visual dynamics in action-chunked policies, drops stale actions on drift, and replans on-the-fly achieving +17.7 success points while keeping the VLA backbone frozen.

A feed-forward model that decomposes unposed images into instance-structured 3D token groups without annotations, enabling unified reconstruction, segmentation, and manipulation.

A proprioceptive contact feedback method leveraging motor current signals as a touch proxy to enable compliant dexterous manipulation without dedicated tactile sensors.

Uses weight-space arithmetic to isolate domain shifts from task knowledge, enabling one-shot VLA policy adaptation to new cameras or embodiments.

Introduces a benchmark showing VLAs lose 20–40 points in commonsense/world knowledge versus source VLMs after robotics fine-tuning, evaluated via action-based answering.

Learns dexterous manipulation policies that explicitly ground actions in generative contact predictions from visuotactile observations, improving robustness on contact-rich tasks.

Introduces multi-axis preference supervision to learn dense, language-conditioned rewards across speed/precision/subtask axes without segmentation; enables compositional generalization and better long-horizon credit assignment than single-reward baselines.

Proposes a general world foundation model leveraging Next-State-Prediction to jointly generate text, images, and embodied actions within a unified framework.

A unified 2B–8B parameter model for robot navigation tasks (VLN, ObjectNav, tracking, autonomous driving) via a configurable observation protocol, with demonstrated zero-shot deployment on real quadruped robots using agentic planners.

Replaces noisy 6-DoF hand poses with relative wrist translation as a shared action space between humans and bimanual robots, enabling scalable skill acquisition from inexpensive video data. This approach outperforms full-pose baselines for robot skill learning.

Large-scale open-source framework for real-world robotic manipulation using behavior cloning, including the ABC-130K dataset with 3,500 hours and 130K+ episodes across 195 tasks. Provides hardware setups, simulators, and training recipes for Diffusion Transformers and Vision-Language-Action models.

Presents alignment techniques to scale robotic manipulation foundation models, building on the Qwen model family for dexterous robot control.

Introduces the first automated system that continuously discovers, evolves, and accumulates reusable sensorimotor skills via evolutionary search over control programs, enabling compounding multi-task, sim-to-real, and cross-embodiment transfer without retraining end-to-end policies.

Embeds both the robot and environment into a shared 4D latent space augmented with forward-kinematics robot points, enabling a vision-language-action model to handle dynamic scenes and long-horizon memory. Outperforms image-only VLA baselines on the BEHAVIOR-1K benchmark for mobile manipulation.

A single-camera sim-to-real framework that uses physically consistent 3D Gaussian Splatting augmentations to achieve zero-shot transfer of dexterous in-hand reorientation policies to an Allegro hand. The approach trains entirely on consumer hardware while maintaining high fidelity to real-world dynamics.

Introduces a high-coverage, conformable robotic skin hardware system designed to improve data collection and policy learning for contact-rich, dexterous manipulation tasks. The system provides rich tactile sensing coverage to enable more capable robot manipulation policies.

A CAT-scan style diagnostic benchmark for robot foundation models that evaluates policies such as π0, π0.5, and Qwen-RobotManip beyond single success rates. The benchmark is designed to distinguish genuine generalization from overfitting to demonstrations in generalist manipulation policies.

A scalable VLA pretraining pipeline that converts unstructured egocentric human videos into robot training data, trains a dexterous hand VLA, and fine-tunes on robot data, achieving strong zero-shot generalization and real-robot dexterous manipulation.