Inner Product Regularization for Imitation Learning
This paper introduces Inner Product Regularization (InneR) to enhance the robustness of robotic policies learned from demonstrations, addressing sensitivity to state perturbations.
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.
This paper introduces Inner Product Regularization (InneR) to enhance the robustness of robotic policies learned from demonstrations, addressing sensitivity to state perturbations.
XBot-Omni is a Thinker--Talker--Actor architecture designed for embodied interaction, integrating multimodal perception and expression in robots.
HEIR is a resource for studying human-robot interactions by pairing human speech, gaze, and egocentric video with robot observations and navigation-manipulation trajectories.
TanCRAFT is a predictive assistance system designed to prevent errors in physical assembly tasks by anticipating user actions and providing timely feedback.
This paper examines how memory-augmented vision-language-action (VLA) policies use time as a proxy for progress, leading to errors in counting tasks.
This paper introduces World In-Context Learning (WICL) to enable robots to learn manipulation skills in context without relying on trained action policies.
This paper explores how generalist AI models can outperform specialist policies in robotic manipulation tasks without task-specific training.
The paper introduces VLA-Tug, a lightweight method for early failure detection and classification in Vision-Language-Action models, improving reliability under distribution shifts.
The paper introduces Latent RL Prior, a two-stage framework that leverages control knowledge from simulation to improve visual policy learning for dexterous manipulation with limited data.
OopsieRecover is a framework that uses semantic interventions to teach VLA models to recover from failures with minimal corrective data.
HALO, a humanoid autonomy system, synthesizes vision-language-motion data from simulated environments to acquire physical skills and transfer them to real-world tasks.
DreamX uses a pretrained video diffusion model to learn a universal action representation that can be transferred across different embodiments and environments.