Advances in Robotic Manipulation and Perception

The field of robotic manipulation and perception is rapidly advancing, with a focus on developing more efficient, generalizable, and robust methods for robots to interact with their environment. Recent research has emphasized the importance of 3D geometry-aware policies, multi-view images, and scene graph-based representations for effective robotic manipulation. Additionally, there is a growing interest in improving the sample efficiency and generalizability of robot policies through techniques such as on-manifold exploration, residual off-policy reinforcement learning, and state-aware guided imitation learning. Notable papers in this area include DIPP, which proposes a discriminative impact point predictor for catching diverse in-flight objects, and GP3, which presents a 3D geometry-aware policy with multi-view images for robotic manipulation. Other noteworthy papers include Compose by Focus, which introduces a scene graph-based representation for compositional generalization, and FUNCanon, which proposes a framework for learning pose-aware action primitives via functional object canonicalization. Overall, these advances have the potential to significantly improve the capabilities of robots in a wide range of applications, from manufacturing and logistics to healthcare and service robotics.

Sources

DIPP: Discriminative Impact Point Predictor for Catching Diverse In-Flight Objects

GP3: A 3D Geometry-Aware Policy with Multi-View Images for Robotic Manipulation

Improving Robotic Manipulation with Efficient Geometry-Aware Vision Encoder

Compose by Focus: Scene Graph-based Atomic Skills

DSPv2: Improved Dense Policy for Effective and Generalizable Whole-body Mobile Manipulation

Do You Need Proprioceptive States in Visuomotor Policies?

N2M: Bridging Navigation and Manipulation by Learning Pose Preference from Rollout

VGGT-DP: Generalizable Robot Control via Vision Foundation Models

FUNCanon: Learning Pose-Aware Action Primitives via Functional Object Canonicalization for Generalizable Robotic Manipulation

SOE: Sample-Efficient Robot Policy Self-Improvement via On-Manifold Exploration

Residual Off-Policy RL for Finetuning Behavior Cloning Policies

SAGE:State-Aware Guided End-to-End Policy for Multi-Stage Sequential Tasks via Hidden Markov Decision Process

Built with on top of