Embodied AI 101

Proposes interleaved vision and language reasoning for robotic manipulation within a VLA framework. Aims to improve instruction following and task performance through integrated multimodal reasoning.

Self-directed play combined with Code-as-Policy for reusable skill acquisition and downstream manipulation tasks.

A unified RL policy for quadrupeds and humanoids that jointly handles force and position control without force sensors, enabling compliant behaviors, force-aware imitation learning, and contact-rich tasks.

Explores imitation learning approaches for multi-robot systems, focusing on policy coupling through sequential asymmetric imitation to enable collaborative robot behaviors.

A humanoid robot 'cerebellum' foundation model trained on 20,000 hours of human motion data that demonstrates scaling laws for robot motion control and enables zero-shot execution of unseen motions on real humanoids.

Combines the Spring-Loaded Inverted Pendulum (SLIP) model with reinforcement learning to achieve agile jumping behaviors in robotic systems.

A 9B model that filters noise from videos, GUI, and embodied data streams, reorganizing them into dense supervision via factual anchors and semantic synthesis; trained with SFT + GRPO across five domains with benchmarks.

Converts 6K+ hours of mixed human/robot egocentric video into robot pseudo-actions via camera-space alignment and reliability-aware loss, achieving 72.8% on RoboCasa and 91.1% on RoboTwin.

A 14B-parameter video world model that converts predicted visual futures into embodiment-agnostic actions via Jacobian inverse-dynamics, enabling zero-shot cross-robot transfer across a Panda arm and 16-DoF hand with open-sourced weights and training code.

A dexterous manipulation foundation model trained on 500k hours of real-world bimanual data that handles deformable objects such as cardboard folding and screw packing, featuring online retry and adaptation capabilities.

Develops an SE(3)-equivariant flow-based visuomotor policy leveraging spherical harmonics for efficient and geometrically consistent robot manipulation.

Introduces a dual-stream transformer architecture inspired by cortical visual processing for learning robotic manipulation policies.

An edge-filtered video streaming agent that evolves skills from memory and runs on smart glasses, reducing API costs by 98%, accompanied by the VisualClawArena benchmark dataset.

A 4B unified architecture for world understanding, generation, and action with hybrid linear attention enabling real-time edge inference across embodiments, outperforming 14B models on embodied benchmarks.

A framework that trains a 51-DoF dexterous hand to open drawers and doors using only physical contact without requiring tactile sensors.

A 4B open-source VLA-style model trained on fewer than 2K simulation trajectories that matches closed frontier systems on real-world manipulation tasks with zero-shot generalization to novel objects and long-horizon behaviors, including failure-recovery demonstrations.

A new geometric action model for robot manipulation policies that focuses on structured action representations to improve policy learning and generalization.

ENPIRE demonstrates fully autonomous physical AutoResearch where Codex agents control a fleet of 8 robots overnight, self-improving through real hardware rollouts on tasks like zip-tie tying and GPU installation, while discovering physical scaling laws with built-in safety harnesses and frozen reward classifiers derived from demonstrations.

An open VLA-style robotics foundation model featuring open weights, open dataset, open action tokenizer, and a depth-reasoning variant; designed to enable community experiments on real robots for manipulation and generalist policies.

Released a massive bimanual robot manipulation dataset with 2,163 hours and 250K+ episodes across 70+ tasks, along with a compatible VLA model for multi-view egocentric teleop. The dataset and model are fully compatible with LeRobot v3.0.

New framework for guided flow-matching policies that improves long-horizon robotic control and sample efficiency.

Introduces flow reversal steering to guide diffusion-based vision-language-action models with high-level VLM reasoning and enables RL directly in the diffusion noise space.

Larger models + more thinking + more context improve performance on some prompts but not others; a learned router enables better performance/latency trade-offs.

RoboGenesis generates 10K+ lab scenes across 16 robot embodiments; LabVLA (Qwen3-VL + DiT flow-matching) achieves 71.1% success on LabUtopia and transfers to real Franka arms.

GPT-style Transformer pretrained on 2 billion motion frames that achieves agile, generalist zero-shot control on a real Unitree G1 humanoid for tasks like soccer, dancing, and digging. Requires no fine-tuning or task-specific adaptation.