Pixel-UMI: Learning Environment-Aware Whole-Body Manipulation
Pixel-UMI is a pipeline for learning environment-aware whole-body policies from sensorized human demonstrations, addressing challenges like embodiment and kinematic differences.
IAEI / RESEARCH & IDEAS
Research briefs on embodied intelligence, robotics and learning through interaction, with links to original sources.
FOLLOW THE RESEARCH
Briefs are automatically prepared from public research abstracts. They do not imply independent verification or association endorsement.
Pixel-UMI is a pipeline for learning environment-aware whole-body policies from sensorized human demonstrations, addressing challenges like embodiment and kinematic differences.
This paper introduces ID-WM, a hierarchical world model that separates state-space reachability from local control, enabling better long-horizon planning.
DAWN is a hierarchical robotic framework using diffusion-based policies and multi-modal sensory integration to automate unfastening tasks, improving recycling and safety.
This tutorial introduces the fundamentals of shape servoing for deformable objects using a data-driven Jacobian-based approach.
The authors propose Gym-Anything for Robotics, a framework using coding agents to create realistic simulations for training embodied AI models, addressing the scalability issue in simulation development.
This research explores advanced in-hand object manipulation using a multi-fingered hand, focusing on complex grasps and safety measures.
ThinkWorld is a VLA framework that uses world dynamics and reinforcement learning to improve robotic manipulation by leveraging human video data.
The paper introduces FLORE, a framework for learning force-reactive humanoid loco-manipulation using robot-free demonstrations and a force-reactive diffusion policy.
RobustRLlib addresses challenges in robust reinforcement learning by providing a unified library and benchmark that enhances algorithm reusability and evaluates empirical capabilities across diverse scenarios.
SceneBloom is a framework that combines visual layout sampling with contract-guided 3D compilation to create realistic and diverse indoor environments for training embodied agents.
This research explores using a VLA model's internal activations to monitor its execution, aiming to reduce the complexity and cost of runtime monitoring compared to existing methods.
DiffDreamer combines world models with differentiable simulation to enhance quadrotor control in tasks like dynamic-gate racing.