Integrating Reinforcement Learning with Foundation Models for Autonomous Robotics: Methods and Perspectives

Foundation models (FMs), large deep learning models pre-trained on vast, unlabeled datasets, exhibit powerful capabilities in understanding complex patterns and generating sophisticated outputs. However, they often struggle to adapt to specific tasks. Reinforcement learning (RL), which allows agents to learn through interaction and feedback, offers a compelling solution. Integrating RL with FMs enables these models to achieve desired outcomes and excel at particular tasks. Additionally, RL can be enhanced by leveraging the reasoning and generalization capabilities of FMs. This synergy is revolutionizing various fields, including robotics. FMs, rich in knowledge and generalization, provide robots with valuable information, while RL facilitates learning and adaptation through real-world interactions. This survey paper comprehensively explores this exciting intersection, examining how these paradigms can be integrated to advance robotic intelligence. We analyze the use of foundation models as action planners, the development of robotics-specific foundation models, and the mutual benefits of combining FMs with RL. Furthermore, we present a taxonomy of integration approaches, including large language models, vision-language models, diffusion models, and transformer-based RL models. We also explore how RL can utilize world representations learned from FMs to enhance robotic task execution. Our survey aims to synthesize current research and highlight key challenges in robotic reasoning and control, particularly in the context of integrating FMs and RL--two rapidly evolving technologies. By doing so, we seek to spark future research and emphasize critical areas that require further investigation to enhance robotics. We provide an updated collection of papers based on our taxonomy, accessible on our open-source project website at: https://github.com/clmoro/Robotics-RL-FMs-Integration.

RoboMorph: In-Context Meta-Learning for Robot Dynamics Modeling

The landscape of Deep Learning has experienced a major shift with the pervasive adoption of Transformer-based architectures, particularly in Natural Language Processing (NLP). Novel avenues for physical applications, such as solving Partial Differential Equations and Image Vision, have been explored. However, in challenging domains like robotics, where high non-linearity poses significant challenges, Transformer-based applications are scarce. While Transformers have been used to provide robots with knowledge about high-level tasks, few efforts have been made to perform system identification. This paper proposes a novel methodology to learn a meta-dynamical model of a high-dimensional physical system, such as the Franka robotic arm, using a Transformer-based architecture without prior knowledge of the system's physical parameters. The objective is to predict quantities of interest (end-effector pose and joint positions) given the torque signals for each joint. This prediction can be useful as a component for Deep Model Predictive Control frameworks in robotics. The meta-model establishes the correlation between torques and positions and predicts the output for the complete trajectory. This work provides empirical evidence of the efficacy of the in-context learning paradigm, suggesting future improvements in learning the dynamics of robotic systems without explicit knowledge of physical parameters. Code, videos, and supplementary materials can be found at project website. See https://sites.google.com/view/robomorph/

Scaling Population-Based Reinforcement Learning with GPU Accelerated Simulation

In recent years, deep reinforcement learning (RL) has shown its effectiveness in solving complex continuous control tasks like locomotion and dexterous manipulation. However, this comes at the cost of an enormous amount of experience required for training, exacerbated by the sensitivity of learning efficiency and the policy performance to hyperparameter selection, which often requires numerous trials of time-consuming experiments. This work introduces a Population-Based Reinforcement Learning (PBRL) approach that exploits a GPU-accelerated physics simulator to enhance the exploration capabilities of RL by concurrently training multiple policies in parallel. The PBRL framework is applied to three state-of-the-art RL algorithms -- PPO, SAC, and DDPG -- dynamically adjusting hyperparameters based on the performance of learning agents. The experiments are performed on four challenging tasks in Isaac Gym -- Anymal Terrain, Shadow Hand, Humanoid, Franka Nut Pick -- by analyzing the effect of population size and mutation mechanisms for hyperparameters. The results show that PBRL agents achieve superior performance, in terms of cumulative reward, compared to non-evolutionary baseline agents. The trained agents are finally deployed in the real world for a Franka Nut Pick task, demonstrating successful sim-to-real transfer. Code and videos of the learned policies are available on our project website.

Combining Sampling- and Gradient-based Planning for Contact-rich Manipulation

Planning over discontinuous dynamics is needed for robotics tasks like contact-rich manipulation, which presents challenges in the numerical stability and speed of planning methods when either neural network or analytical models are used. On the one hand, sampling-based planners require higher sample complexity in high-dimensional problems and cannot describe safety constraints such as force limits. On the other hand, gradient-based solvers can suffer from local optima and convergence issues when the Hessian is poorly conditioned. We propose a planning method with both sampling- and gradient-based elements, using the Cross-entropy Method to initialize a gradient-based solver, providing better search over local minima and the ability to handle explicit constraints. We show the approach allows smooth, stable contact-rich planning for an impedance-controlled robot making contact with a stiff environment, benchmarking against gradient-only MPC and CEM.

Imitation Learning-based Visual Servoing for Tracking Moving Objects

In everyday life collaboration tasks between human operators and robots, the former necessitate simple ways for programming new skills, the latter have to show adaptive capabilities to cope with environmental changes. The joint use of visual servoing and imitation learning allows us to pursue the objective of realizing friendly robotic interfaces that (i) are able to adapt to the environment thanks to the use of visual perception and (ii) avoid explicit programming thanks to the emulation of previous demonstrations. This work aims to exploit imitation learning for the visual servoing paradigm to address the specific problem of tracking moving objects. In particular, we show that it is possible to infer from data the compensation term required for realizing the tracking controller, avoiding the explicit implementation of estimators or observers. The effectiveness of the proposed method has been validated through simulations with a robotic manipulator.

Predicting human motion intention for pHRI assistive control

This work addresses human intention identification during physical Human-Robot Interaction (pHRI) tasks to include this information in an assistive controller. To this purpose, human intention is defined as the desired trajectory that the human wants to follow over a finite rolling prediction horizon so that the robot can assist in pursuing it. This work investigates a Recurrent Neural Network (RNN), specifically, Long-Short Term Memory (LSTM) cascaded with a Fully Connected layer. In particular, we propose an iterative training procedure to adapt the model. Such an iterative procedure is powerful in reducing the prediction error. Still, it has the drawback that it is time-consuming and does not generalize to different users or different co-manipulated objects. To overcome this issue, Transfer Learning (TL) adapts the pre-trained model to new trajectories, users, and co-manipulated objects by freezing the LSTM layer and fine-tuning the last FC layer, which makes the procedure faster. Experiments show that the iterative procedure adapts the model and reduces prediction error. Experiments also show that TL adapts to different users and to the co-manipulation of a large object. Finally, to check the utility of adopting the proposed method, we compare the proposed controller enhanced by the intention prediction with the other two standard controllers of pHRI.

Teaching contact-rich tasks from visual demonstrations by constraint extraction

Contact-rich manipulation involves kinematic constraints on the task motion, typically with discrete transitions between these constraints during the task. Allowing the robot to detect and reason about these contact constraints can support robust and dynamic manipulation, but how can these contact models be efficiently learned? Purely visual observations are an attractive data source, allowing passive task demonstrations with unmodified objects. Existing approaches for vision-only learning from demonstration are effective in pick-and-place applications and planar tasks. Nevertheless, accuracy/occlusions and unobserved task dynamics can limit their robustness in contact-rich manipulation. To use visual demonstrations for contact-rich robotic tasks, we consider the demonstration of pose trajectories with transitions between holonomic kinematic constraints, first clustering the trajectories into discrete contact modes, then fitting kinematic constraints per each mode. The fit constraints are then used to (i) detect contact online with force/torque measurements and (ii) plan the robot policy with respect to the active constraint. We demonstrate the approach with real experiments, on cabling and rake tasks, showing the approach gives robust manipulation through contact transitions.

Differentiable Environment Primitives for Contact State Estimation

In contact-rich manipulation, the robot dynamics are coupled with an environment that has application-specific dynamic properties (stiffness, inertia) and geometry (contact normal). Knowledge of these environmental parameters can improve control and monitoring, but they are often unobserved and may vary, either online or between task instances. Observers, such as the extended Kalman filter, can be used to estimate these parameters, but such model-based techniques can require too much engineering work to scale up to complex environments, such as multi-point contact. To accelerate environment modeling, we propose environment primitives: parameterized environment dynamics that can be connected in parallel and are expressed in an automatic differentiation framework. This simplifies offline gradient-based optimization to fit model parameters and linearization of the coupled dynamics for an observer. This method is implemented for stiffness contact models, allowing the fitting of contact geometry and stiffness offline or their online estimation by an extended Kalman filter. This method is applied to a collaborative robot, estimating external force, contact stiffness, and contact geometry from the motor position and current. The estimates of external force and stiffness are compared with a momentum observer and direct force measurements.

Experience in Engineering Complex Systems: Active Preference Learning with Multiple Outcomes and Certainty Levels

Black-box optimization refers to the optimization problem whose objective function and/or constraint sets are either unknown, inaccessible, or non-existent. In many applications, especially with the involvement of humans, the only way to access the optimization problem is through performing physical experiments with the available outcomes being the preference of one candidate with respect to one or many others. Accordingly, the algorithm so-called Active Preference Learning has been developed to exploit this specific information in constructing a surrogate function based on standard radial basis functions, and then forming an easy-to-solve acquisition function which repetitively suggests new decision vectors to search for the optimal solution. Based on this idea, our approach aims to extend the algorithm in such a way that can exploit further information effectively, which can be obtained in reality such as: 5-point Likert type scale for the outcomes of the preference query (i.e., the preference can be described in not only "this is better than that" but also "this is much better than that" level), or multiple outcomes for a single preference query with possible additive information on how certain the outcomes are. The validation of the proposed algorithm is done through some standard benchmark functions, showing a promising improvement with respect to the state-of-the-art algorithm.

Environment-based Assistance Modulation for a Hip Exosuit via Computer Vision

Just like in humans vision plays a fundamental role in guiding adaptive locomotion, when designing the control strategy for a walking assistive technology, Computer Vision may bring substantial improvements when performing an environment-based assistance modulation. In this work, we developed a hip exosuit controller able to distinguish among three different walking terrains through the use of an RGB camera and to adapt the assistance accordingly. The system was tested with seven healthy participants walking throughout an overground path comprising of staircases and level ground. Subjects performed the task with the exosuit disabled (Exo Off), constant assistance profile (Vision Off ), and with assistance modulation (Vision On). Our results showed that the controller was able to promptly classify in real-time the path in front of the user with an overall accuracy per class above the 85%, and to perform assistance modulation accordingly. Evaluation related to the effects on the user showed that Vision On was able to outperform the other two conditions: we obtained significantly higher metabolic savings than Exo Off, with a peak of about -20% when climbing up the staircase and about -16% in the overall path, and than Vision Off when ascending or descending stairs. Such advancements in the field may yield to a step forward for the exploitation of lightweight walking assistive technologies in real-life scenarios.