RLingua: Improving Reinforcement Learning Sample Efficiency in Robotic Manipulations With Large Language Models

Reinforcement learning (RL) has demonstrated its capability in solving various tasks but is notorious for its low sample efficiency. In this paper, we propose RLingua, a framework that can leverage the internal knowledge of large language models (LLMs) to reduce the sample complexity of RL in robotic manipulations. To this end, we first present a method for extracting the prior knowledge of LLMs by prompt engineering so that a preliminary rule-based robot controller for a specific task can be generated in a user-friendly manner. Despite being imperfect, the LLM-generated robot controller is utilized to produce action samples during rollouts with a decaying probability, thereby improving RL's sample efficiency. We employ TD3, the widely-used RL baseline method, and modify the actor loss to regularize the policy learning towards the LLM-generated controller. RLingua also provides a novel method of improving the imperfect LLM-generated robot controllers by RL. We demonstrate that RLingua can significantly reduce the sample complexity of TD3 in four robot tasks of panda_gym and achieve high success rates in 12 sampled sparsely rewarded robot tasks in RLBench, where the standard TD3 fails. Additionally, We validated RLingua's effectiveness in real-world robot experiments through Sim2Real, demonstrating that the learned policies are effectively transferable to real robot tasks. Further details about our work are available at our project website https://rlingua.github.io.

VIHE: Virtual In-Hand Eye Transformer for 3D Robotic Manipulation

In this work, we introduce the Virtual In-Hand Eye Transformer (VIHE), a novel method designed to enhance 3D manipulation capabilities through action-aware view rendering. VIHE autoregressively refines actions in multiple stages by conditioning on rendered views posed from action predictions in the earlier stages. These virtual in-hand views provide a strong inductive bias for effectively recognizing the correct pose for the hand, especially for challenging high-precision tasks such as peg insertion. On 18 manipulation tasks in RLBench simulated environments, VIHE achieves a new state-of-the-art, with a 12% absolute improvement, increasing from 65% to 77% over the existing state-of-the-art model using 100 demonstrations per task. In real-world scenarios, VIHE can learn manipulation tasks with just a handful of demonstrations, highlighting its practical utility. Videos and code implementation can be found at our project site: https://vihe-3d.github.io.

Reasoning Grasping via Multimodal Large Language Model

Despite significant progress in robotic systems for operation within human-centric environments, existing models still heavily rely on explicit human commands to identify and manipulate specific objects. This limits their effectiveness in environments where understanding and acting on implicit human intentions are crucial. In this study, we introduce a novel task: reasoning grasping, where robots need to generate grasp poses based on indirect verbal instructions or intentions. To accomplish this, we propose an end-to-end reasoning grasping model that integrates a multi-modal Large Language Model (LLM) with a vision-based robotic grasping framework. In addition, we present the first reasoning grasping benchmark dataset generated from the GraspNet-1 billion, incorporating implicit instructions for object-level and part-level grasping, and this dataset will soon be available for public access. Our results show that directly integrating CLIP or LLaVA with the grasp detection model performs poorly on the challenging reasoning grasping tasks, while our proposed model demonstrates significantly enhanced performance both in the reasoning grasping benchmark and real-world experiments.

Learning Excavation of Rigid Objects with Offline Reinforcement Learning

Autonomous excavation is a challenging task. The unknown contact dynamics between the excavator bucket and the terrain could easily result in large contact forces and jamming problems during excavation. Traditional model-based methods struggle to handle such problems due to complex dynamic modeling. In this paper, we formulate the excavation skills with three novel manipulation primitives. We propose to learn the manipulation primitives with offline reinforcement learning (RL) to avoid large amounts of online robot interactions. The proposed method can learn efficient penetration skills from sub-optimal demonstrations, which contain sub-trajectories that can be ``stitched" together to formulate an optimal trajectory without causing jamming. We evaluate the proposed method with extensive experiments on excavating a variety of rigid objects and demonstrate that the learned policy outperforms the demonstrations. We also show that the learned policy can quickly adapt to unseen and challenging fragmented rocks with online fine-tuning.

GOATS: Goal Sampling Adaptation for Scooping with Curriculum Reinforcement Learning

In this work, we first formulate the problem of goal-conditioned robotic water scooping with reinforcement learning. This task is challenging due to the complex dynamics of fluid and multi-modal goal-reaching. The policy is required to achieve both position goals and water amount goals, which leads to a large convoluted goal state space. To address these challenges, we introduce Goal Sampling Adaptation for Scooping (GOATS), a curriculum reinforcement learning method that can learn an effective and generalizable policy for robot scooping tasks. Specifically, we use a goal-factorized reward formulation and interpolate position goal distributions and amount goal distributions to create curriculum through the learning process. As a result, our proposed method can outperform the baselines in simulation and achieves 5.46% and 8.71% amount errors on bowl scooping and bucket scooping tasks, respectively, under 1000 variations of initial water states in the tank and a large goal state space. Besides being effective in simulation environments, our method can efficiently generalize to noisy real-robot water-scooping scenarios with different physical configurations and unseen settings, demonstrating superior efficacy and generalizability. The videos of this work are available on our project page: https://sites.google.com/view/goatscooping.

Offline-Online Learning of Deformation Model for Cable Manipulation with Graph Neural Networks

Manipulating deformable linear objects by robots has a wide range of applications, e.g., manufacturing and medical surgery. To complete such tasks, an accurate dynamics model for predicting the deformation is critical for robust control. In this work, we deal with this challenge by proposing a hybrid offline-online method to learn the dynamics of cables in a robust and data-efficient manner. In the offline phase, we adopt Graph Neural Network (GNN) to learn the deformation dynamics purely from the simulation data. Then a linear residual model is learned in real-time to bridge the sim-to-real gap. The learned model is then utilized as the dynamics constraint of a trust region based Model Predictive Controller (MPC) to calculate the optimal robot movements. The online learning and MPC run in a closed-loop manner to robustly accomplish the task. Finally, comparative results with existing methods are provided to quantitatively show the effectiveness and robustness.

Learning Insertion Primitives with Discrete-Continuous Hybrid Action Space for Robotic Assembly Tasks

This paper introduces a discrete-continuous action space to learn insertion primitives for robotic assembly tasks. Primitive is a sequence of elementary actions with certain exit conditions, such as "pushing down the peg until contact". Since the primitive is an abstraction of robot control commands and encodes human prior knowledge, it reduces the exploration difficulty and yields better learning efficiency. In this paper, we learn robot assembly skills via primitives. Specifically, we formulate insertion primitives as parameterized actions: hybrid actions consisting of discrete primitive types and continuous primitive parameters. Compared with the previous work using a set of discretized parameters for each primitive, the agent in our method can freely choose primitive parameters from a continuous space, which is more flexible and efficient. To learn these insertion primitives, we propose Twin-Smoothed Multi-pass Deep Q-Network (TS-MP-DQN), an advanced version of MP-DQN with twin Q-network to reduce the Q-value over-estimation. Extensive experiments are conducted in the simulation and real world for validation. From experiment results, our approach achieves higher success rates than three baselines: MP-DQN with parameterized actions, primitives with discrete parameters, and continuous velocity control. Furthermore, learned primitives are robust to sim-to-real transfer and can generalize to challenging assembly tasks such as tight round peg-hole and complex shaped electric connectors with promising success rates. Experiment videos are available at https://msc.berkeley.edu/research/insertion-primitives.html.

Trajectory Optimization for Manipulation of Deformable Objects: Assembly of Belt Drive Units

This paper presents a novel trajectory optimization formulation to solve the robotic assembly of the belt drive unit. Robotic manipulations involving contacts and deformable objects are challenging in both dynamic modeling and trajectory planning. For modeling, variations in the belt tension and contact forces between the belt and the pulley could dramatically change the system dynamics. For trajectory planning, it is computationally expensive to plan trajectories for such hybrid dynamical systems as it usually requires planning for discrete modes separately. In this work, we formulate the belt drive unit assembly task as a trajectory optimization problem with complementarity constraints to avoid explicitly imposing contact mode sequences. The problem is solved as a mathematical program with complementarity constraints (MPCC) to obtain feasible and efficient assembly trajectories. We validate the proposed method both in simulations with a physics engine and in real-world experiments with a robotic manipulator.