Visual Imitation Learning of Non-Prehensile Manipulation Tasks with Dynamics-Supervised Models

Unlike quasi-static robotic manipulation tasks like pick-and-place, dynamic tasks such as non-prehensile manipulation pose greater challenges, especially for vision-based control. Successful control requires the extraction of features relevant to the target task. In visual imitation learning settings, these features can be learnt by backpropagating the policy loss through the vision backbone. Yet, this approach tends to learn task-specific features with limited generalizability. Alternatively, learning world models can realize more generalizable vision backbones. Utilizing the learnt features, task-specific policies are subsequently trained. Commonly, these models are trained solely to predict the next RGB state from the current state and action taken. But only-RGB prediction might not fully-capture the task-relevant dynamics. In this work, we hypothesize that direct supervision of target dynamic states (Dynamics Mapping) can learn better dynamics-informed world models. Beside the next RGB reconstruction, the world model is also trained to directly predict position, velocity, and acceleration of environment rigid bodies. To verify our hypothesis, we designed a non-prehensile 2D environment tailored to two tasks: "Balance-Reaching" and "Bin-Dropping". When trained on the first task, dynamics mapping enhanced the task performance under different training configurations (Decoupled, Joint, End-to-End) and policy architectures (Feedforward, Recurrent). Notably, its most significant impact was for world model pretraining boosting the success rate from 21% to 85%. Although frozen dynamics-informed world models could generalize well to a task with in-domain dynamics, but poorly to a one with out-of-domain dynamics.

Achieving Faster and More Accurate Operation of Deep Predictive Learning

Achieving both high speed and precision in robot operations is a significant challenge for social implementation. While factory robots excel at predefined tasks, they struggle with environment-specific actions like cleaning and cooking. Deep learning research aims to address this by enabling robots to autonomously execute behaviors through end-to-end learning with sensor data. RT-1 and ACT are notable examples that have expanded robots' capabilities. However, issues with model inference speed and hand position accuracy persist. High-quality training data and fast, stable inference mechanisms are essential to overcome these challenges. This paper proposes a motion generation model for high-speed, high-precision tasks, exemplified by the sports stacking task. By teaching motions slowly and inferring at high speeds, the model achieved a 94% success rate in stacking cups with a real robot.

Dual-arm Motion Generation for Repositioning Care based on Deep Predictive Learning with Somatosensory Attention Mechanism

A versatile robot working in a domestic environment based on a deep neural network (DNN) is currently attracting attention. One of the roles expected for domestic robots is caregiving for a human. In particular, we focus on repositioning care because repositioning plays a fundamental role in supporting the health and quality of life of individuals with limited mobility. However, generating motions of the repositioning care, avoiding applying force to non-target parts and applying appropriate force to target parts, remains challenging. In this study, we proposed a DNN-based architecture using visual and somatosensory attention mechanisms that can generate dual-arm repositioning motions which involve different sequential policies of interaction force; contact-less reaching and contact-based assisting motions. We used the humanoid robot Dry-AIREC, which features the capability to adjust joint impedance dynamically. In the experiment, the repositioning assistance from the supine position to the sitting position was conducted by Dry-AIREC. The trained model, utilizing the proposed architecture, successfully guided the robot's hand to the back of the mannequin without excessive contact force on the mannequin and provided adequate support and appropriate contact for postural adjustment.

Sensorimotor Attention and Language-based Regressions in Shared Latent Variables for Integrating Robot Motion Learning and LLM

In recent years, studies have been actively conducted on combining large language models (LLM) and robotics; however, most have not considered end-to-end feedback in the robot-motion generation phase. The prediction of deep neural networks must contain errors, it is required to update the trained model to correspond to the real environment to generate robot motion adaptively. This study proposes an integration method that connects the robot-motion learning model and LLM using shared latent variables. When generating robot motion, the proposed method updates shared parameters based on prediction errors from both sensorimotor attention points and task language instructions given to the robot. This allows the model to search for latent parameters appropriate for the robot task efficiently. Through simulator experiments on multiple robot tasks, we demonstrated the effectiveness of our proposed method from two perspectives: position generalization and language instruction generalization abilities.

A Peg-in-hole Task Strategy for Holes in Concrete

A method that enables an industrial robot to accomplish the peg-in-hole task for holes in concrete is proposed. The proposed method involves slightly detaching the peg from the wall, when moving between search positions, to avoid the negative influence of the concrete's high friction coefficient. It uses a deep neural network (DNN), trained via reinforcement learning, to effectively find holes with variable shape and surface finish (due to the brittle nature of concrete) without analytical modeling or control parameter tuning. The method uses displacement of the peg toward the wall surface, in addition to force and torque, as one of the inputs of the DNN. Since the displacement increases as the peg gets closer to the hole (due to the chamfered shape of holes in concrete), it is a useful parameter for inputting in the DNN. The proposed method was evaluated by training the DNN on a hole 500 times and attempting to find 12 unknown holes. The results of the evaluation show the DNN enabled a robot to find the unknown holes with average success rate of 96.1% and average execution time of 12.5 seconds. Additional evaluations with random initial positions and a different type of peg demonstrate the trained DNN can generalize well to different conditions. Analyses of the influence of the peg displacement input showed the success rate of the DNN is increased by utilizing this parameter. These results validate the proposed method in terms of its effectiveness and applicability to the construction industry.

Visual Spatial Attention and Proprioceptive Data-Driven Reinforcement Learning for Robust Peg-in-Hole Task Under Variable Conditions

Anchor-bolt insertion is a peg-in-hole task performed in the construction field for holes in concrete. Efforts have been made to automate this task, but the variable lighting and hole surface conditions, as well as the requirements for short setup and task execution time make the automation challenging. In this study, we introduce a vision and proprioceptive data-driven robot control model for this task that is robust to challenging lighting and hole surface conditions. This model consists of a spatial attention point network (SAP) and a deep reinforcement learning (DRL) policy that are trained jointly end-to-end to control the robot. The model is trained in an offline manner, with a sample-efficient framework designed to reduce training time and minimize the reality gap when transferring the model to the physical world. Through evaluations with an industrial robot performing the task in 12 unknown holes, starting from 16 different initial positions, and under three different lighting conditions (two with misleading shadows), we demonstrate that SAP can generate relevant attention points of the image even in challenging lighting conditions. We also show that the proposed model enables task execution with higher success rate and shorter task completion time than various baselines. Due to the proposed model's high effectiveness even in severe lighting, initial positions, and hole conditions, and the offline training framework's high sample-efficiency and short training time, this approach can be easily applied to construction.

Realtime Motion Generation with Active Perception Using Attention Mechanism for Cooking Robot

To support humans in their daily lives, robots are required to autonomously learn, adapt to objects and environments, and perform the appropriate actions. We tackled on the task of cooking scrambled eggs using real ingredients, in which the robot needs to perceive the states of the egg and adjust stirring movement in real time, while the egg is heated and the state changes continuously. In previous works, handling changing objects was found to be challenging because sensory information includes dynamical, both important or noisy information, and the modality which should be focused on changes every time, making it difficult to realize both perception and motion generation in real time. We propose a predictive recurrent neural network with an attention mechanism that can weigh the sensor input, distinguishing how important and reliable each modality is, that realize quick and efficient perception and motion generation. The model is trained with learning from the demonstration, and allows the robot to acquire human-like skills. We validated the proposed technique using the robot, Dry-AIREC, and with our learning model, it could perform cooking eggs with unknown ingredients. The robot could change the method of stirring and direction depending on the status of the egg, as in the beginning it stirs in the whole pot, then subsequently, after the egg started being heated, it starts flipping and splitting motion targeting specific areas, although we did not explicitly indicate them.

Real-time Motion Generation and Data Augmentation for Grasping Moving Objects with Dynamic Speed and Position Changes

While deep learning enables real robots to perform complex tasks had been difficult to implement in the past, the challenge is the enormous amount of trial-and-error and motion teaching in a real environment. The manipulation of moving objects, due to their dynamic properties, requires learning a wide range of factors such as the object's position, movement speed, and grasping timing. We propose a data augmentation method for enabling a robot to grasp moving objects with different speeds and grasping timings at low cost. Specifically, the robot is taught to grasp an object moving at low speed using teleoperation, and multiple data with different speeds and grasping timings are generated by down-sampling and padding the robot sensor data in the time-series direction. By learning multiple sensor data in a time series, the robot can generate motions while adjusting the grasping timing for unlearned movement speeds and sudden speed changes. We have shown using a real robot that this data augmentation method facilitates learning the relationship between object position and velocity and enables the robot to perform robust grasping motions for unlearned positions and objects with dynamically changing positions and velocities.

Interactively Robot Action Planning with Uncertainty Analysis and Active Questioning by Large Language Model

The application of the Large Language Model (LLM) to robot action planning has been actively studied. The instructions given to the LLM by natural language may include ambiguity and lack of information depending on the task context. It is possible to adjust the output of LLM by making the instruction input more detailed; however, the design cost is high. In this paper, we propose the interactive robot action planning method that allows the LLM to analyze and gather missing information by asking questions to humans. The method can minimize the design cost of generating precise robot instructions. We demonstrated the effectiveness of our method through concrete examples in cooking tasks. However, our experiments also revealed challenges in robot action planning with LLM, such as asking unimportant questions and assuming crucial information without asking. Shedding light on these issues provides valuable insights for future research on utilizing LLM for robotics.

Deep Predictive Learning : Motion Learning Concept inspired by Cognitive Robotics

A deep learning-based approach can generalize model performance while reducing feature design costs by learning end-to-end environment recognition and motion generation. However, the process incurs huge training data collection costs and time and human resources for trial-and-error when involving physical contact with robots. Therefore, we propose ``deep predictive learning,'' a motion learning concept that assumes imperfections in the predictive model and minimizes the prediction error with the real-world situation. Deep predictive learning is inspired by the ``free energy principle and predictive coding theory,'' which explains how living organisms behave to minimize the prediction error between the real world and the brain. Robots predict near-future situations based on sensorimotor information and generate motions that minimize the gap with reality. The robot can flexibly perform tasks in unlearned situations by adjusting its motion in real-time while considering the gap between learning and reality. This paper describes the concept of deep predictive learning, its implementation, and examples of its application to real robots. The code and document are available at https: //ogata-lab.github.io/eipl-docs