Balancing Data for Mega-Scale RL in Robot Control
This paper addresses the exploration bottleneck in mega-scale reinforcement learning for robot control by introducing Success Guided Sampling (SGS) to optimize task configuration sampling.
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 addresses the exploration bottleneck in mega-scale reinforcement learning for robot control by introducing Success Guided Sampling (SGS) to optimize task configuration sampling.
DreamTrue is a multi-view, cross-embodiment robot world model designed for accurate video prediction. It addresses calibration issues and interaction biases through offline geometric calibration and counterfactual post-training.
RoboPrompt is a lightweight system enabling intuitive human guidance of robot policies using sparse inputs like drawn traces and directional cues.
QF3 is an online off-policy RL algorithm that enhances flow policies using filtered Q-gradients, enabling fast training and transfer to hardware.
InterMimicGen is a self-evolving framework that enhances humanoid robot motion imitation by iteratively improving motion data and tracking policies.
EyeRobot 2.0 uses active gaze to enable precise bimanual manipulation with a single stereo camera, improving performance over traditional methods without wrist cameras.
This research explores how a helper robot can infer physical constraints of a partner robot through observation and use this information to coordinate on new tasks.
InterEvolve explores evolving reward programs at test time to enable humanoid robots to perform tasks they were not trained for, leveraging their existing skills and learning from attempts without retraining.
The RPG framework enables robots to autonomously improve their execution without updating model weights by reconstructing manipulation capabilities, practicing in simulation, and refining skills based on feedback.
This paper introduces a continual learning framework for 6-DoF grasp synthesis using memory and demonstrations, evaluated in simulation and real-world experiments.
This paper introduces skill-space shooting, a method enabling robots to improve their policies autonomously by leveraging reusable skills.
This research explores the importance of spatial information in vision-language-action (VLA) pruning strategies for robotic tasks, proposing a method called GeoScaffold that maintains spatial structures while reducing token usage.