robot-learning

Robotics teams are rebuilding the data stack from scratch to overcome the 'data layer tax' that slows down iteration and scaling in robot learning, as existing infrastructure doesn't handle multi-rate and multimodal data.

EgoSteer presents a full-stack system that pre-trains a vision-language-action model from egocentric human videos for steerable dexterous manipulation, enabling robust generalization across 40+ diverse tasks with 75%+ success on complex long-horizon tasks.

PoLAR introduces a geometrically structured latent action representation in hyperbolic space that separates transition extent from mode, improving robotic policy learning performance.

This paper finds that egocentric human video, when processed with a filtering and labeling pipeline, can outperform teleoperated real-robot data for pretraining embodied foundation models, achieving lower validation loss and higher success rates on real-robot tasks.

Introduces Playful Agentic Robot Learning, where embodied coding agents use self-directed play to learn reusable skills, improving downstream task performance without additional training. The proposed RATs system achieves significant gains over baselines in simulation and real-world transfer.

Geometric Action Model repurposes a geometric foundation model for robot policy learning, achieving 85.5% on LIBERO-Plus with 6.9 ms inference, 55× faster than baselines.

This paper proposes Hierarchical Advantage-Weighted Behavior Cloning (HABC) for fine-tuning Vision-Language-Action (VLA) policies using online reinforcement learning with sparse binary episode outcomes. HABC separates viability and efficiency objectives via adaptive critic heads and intervention-aware credit assignment, significantly improving success rates on contact-rich bimanual manipulation tasks.

RL4IL introduces a reinforcement learning-guided retrieval method that uses soft fusion over frozen demonstration libraries to handle missing sensor modalities in robotic imitation learning at inference time, achieving high success rates under complete camera dropout.

Lucky Robots announces Lucky Engine, the first game engine purpose-built for robotics, enabling infinite data generation for robotic AI training through realistic simulation and deployment.

HyVLA-0.5 is an end-to-end robotic learning system that integrates data collection, model design, pre-training, fine-tuning, and reinforcement learning for real-world deployment.

μ_0 is a scalable world model that predicts smooth 3D trajectories for interaction points, enabling embodiment-agnostic robot learning without action labels by using a TraceExtract system for supervision.

ETH Zurich has open-sourced their entire 2026 robot learning course, including slides, lecture recordings, coding assignments, and a GitHub repository, covering topics from imitation learning to foundation models for robotics, with guest lectures from industry leaders.

This position paper argues that advancing robot intelligence requires integrating unstructured behavioral data through specialized interfaces for labeling, embodiment mapping, world modeling, and reward inference, rather than relying solely on scaling Vision-Language-Action (VLA) models and world models.

Hugging Face launched LeLab, a graphical user interface for LeRobot that enables training AI robots without command-line interaction, featuring zero-terminal setup, data collection, and one-click GPU training via Hugging Face Jobs.

GRAIL generates diverse humanoid manipulation and locomotion data using 3D assets and video foundation models, enabling effective sim-to-real transfer for humanoid robot control with high real-world success rates.

This paper introduces Parameterized Diffusion Policy (PDP), a framework that makes diffusion policies controllable by conditioning on low-dimensional latent parameters, enabling smooth behavior interpolation and adaptation without retraining. It demonstrates improved performance on complex multimodal robot tasks in simulation and real-world experiments.

WALL-WM advances video-action learning by using semantic events as learning units instead of fixed action chunks, enabling more flexible and scalable vision-language-action training and inference.

Proposes the Agentic-VLA framework, introducing agents into the VLA loop, enabling the vision-language-action model to self-evolve and surpass existing leading embodied models on all metrics.

This paper proposes a thinking-learning interaction model for autonomous robots, enabling them to adaptively discover new features, expand output categories, update learning models, and reconstruct action routines in open environments. Experimental results demonstrate significant improvements in recognition accuracy, category formation, and action efficiency.

Hugging Face released LeRobotHF, an open-source humanoid robot platform with hardware, simulation, and training tools, costing roughly $2,500 to build.