Predicting Motion Plans for Articulating Everyday Objects

概述

accepted：ICRA2023
项目主页
文章解读参考：
    

论文解读

工作动机

从环境约束任务约束的角度(under environmental and task constraints)阐述传统运动规划算法(classical motion planning approaches)在家庭环境移动操作(Mobile manipulation tasks such as opening a door, pulling open a drawer, or lifting a toilet seat)种所面临的挑战。要求机器人给环境施加合适的力，同时推理出可行的轨迹(applying forces onto the environment,while maintaining relevant contact,reasoning about the feasibility of the entire trajectory,轨迹既要满足物品的关节约束，也要保证机器人不与物体本身或周围障碍物碰撞)。文章所采用的解决思路是将其视作一个学习问题(cast it as a learning problem)，借鉴过去的经验(leverage past experience)解决类似的任务规划。This paper develops datasets and techniques for learning models that can predict motion plans for such constrained motion planning problems with low sensing and planning costs.

因为约束的存在，满足条件的轨迹很可能是一个零测度集(measure zero set),这就导致了基于采样的算法失效了，或者必须投射到约束流形中求解(project states to the constraint manifold)。

方法架构

构建了ArtObjSim一个轻量级的运动学模拟器(放置在真实场景中的日常物品,from HM3D dataset [4],10 object categories in 97 scenes).

不直接预测运动规划，而是预测一个可以解码成运动规划的策略(predict a strategy which is decoded into a complete motion plan that adheres to the task constraints at hand),同时通过逆运动学问题求解路径点(Sequential Inverse Kinematics，SeqIK)。所谓策略，是指第一个时间步的 θ 0 \theta_0