Multi-task learning solves multiple correlated tasks. However, conflicts may exist between them. In such circumstances, a single solution can rarely optimize all the tasks, leading to performance trade-offs. To arrive at a set of optimized yet well-distributed models that collectively embody different trade-offs in one algorithmic pass, this paper proposes to view Pareto multi-task learning through the lens of multi-task optimization. Multi-task learning is first cast as a multi-objective optimization problem, which is then decomposed into a diverse set of unconstrained scalar-valued subproblems. These subproblems are solved jointly using a novel multi-task gradient descent method, whose uniqueness lies in the iterative transfer of model parameters among the subproblems during the course of optimization. A theorem proving faster convergence through the inclusion of such transfers is presented. We investigate the proposed multi-task learning with multi-task optimization for solving various problem settings including image classification, scene understanding, and multi-target regression. Comprehensive experiments confirm that the proposed method significantly advances the state-of-the-art in discovering sets of Pareto-optimized models. Notably, on the large image dataset we tested on, namely NYUv2, the hypervolume convergence achieved by our method was found to be nearly two times faster than the next-best among the state-of-the-art.
The creation of large, diverse, high-quality robot manipulation datasets is an important stepping stone on the path toward more capable and robust robotic manipulation policies. However, creating such datasets is challenging: collecting robot manipulation data in diverse environments poses logistical and safety challenges and requires substantial investments in hardware and human labour. As a result, even the most general robot manipulation policies today are mostly trained on data collected in a small number of environments with limited scene and task diversity. In this work, we introduce DROID (Distributed Robot Interaction Dataset), a diverse robot manipulation dataset with 76k demonstration trajectories or 350 hours of interaction data, collected across 564 scenes and 84 tasks by 50 data collectors in North America, Asia, and Europe over the course of 12 months. We demonstrate that training with DROID leads to policies with higher performance and improved generalization ability. We open source the full dataset, policy learning code, and a detailed guide for reproducing our robot hardware setup.
Intelligent agents must be generalists - showing the ability to quickly adapt and generalize to varying tasks. Within the framework of reinforcement learning (RL), model-based RL algorithms learn a task-agnostic dynamics model of the world, in principle allowing them to generalize to arbitrary rewards. However, one-step models naturally suffer from compounding errors, making them ineffective for problems with long horizons and large state spaces. In this work, we propose a novel class of models - generalized occupancy models (GOMs) - that retain the generality of model-based RL while avoiding compounding error. The key idea behind GOMs is to model the distribution of all possible long-term outcomes from a given state under the coverage of a stationary dataset, along with a policy that realizes a particular outcome from the given state. These models can then quickly be used to select the optimal action for arbitrary new tasks, without having to redo policy optimization. By directly modeling long-term outcomes, GOMs avoid compounding error while retaining generality across arbitrary reward functions. We provide a practical instantiation of GOMs using diffusion models and show its efficacy as a new class of transferable models, both theoretically and empirically across a variety of simulated robotics problems. Videos and code at https://weirdlabuw.github.io/gom/.
Imitation learning methods need significant human supervision to learn policies robust to changes in object poses, physical disturbances, and visual distractors. Reinforcement learning, on the other hand, can explore the environment autonomously to learn robust behaviors but may require impractical amounts of unsafe real-world data collection. To learn performant, robust policies without the burden of unsafe real-world data collection or extensive human supervision, we propose RialTo, a system for robustifying real-world imitation learning policies via reinforcement learning in "digital twin" simulation environments constructed on the fly from small amounts of real-world data. To enable this real-to-sim-to-real pipeline, RialTo proposes an easy-to-use interface for quickly scanning and constructing digital twins of real-world environments. We also introduce a novel "inverse distillation" procedure for bringing real-world demonstrations into simulated environments for efficient fine-tuning, with minimal human intervention and engineering required. We evaluate RialTo across a variety of robotic manipulation problems in the real world, such as robustly stacking dishes on a rack, placing books on a shelf, and six other tasks. RialTo increases (over 67%) in policy robustness without requiring extensive human data collection. Project website and videos at https://real-to-sim-to-real.github.io/RialTo/
In recent years, significant progress has been made in the field of robotic reinforcement learning (RL), enabling methods that handle complex image observations, train in the real world, and incorporate auxiliary data, such as demonstrations and prior experience. However, despite these advances, robotic RL remains hard to use. It is acknowledged among practitioners that the particular implementation details of these algorithms are often just as important (if not more so) for performance as the choice of algorithm. We posit that a significant challenge to widespread adoption of robotic RL, as well as further development of robotic RL methods, is the comparative inaccessibility of such methods. To address this challenge, we developed a carefully implemented library containing a sample efficient off-policy deep RL method, together with methods for computing rewards and resetting the environment, a high-quality controller for a widely-adopted robot, and a number of challenging example tasks. We provide this library as a resource for the community, describe its design choices, and present experimental results. Perhaps surprisingly, we find that our implementation can achieve very efficient learning, acquiring policies for PCB board assembly, cable routing, and object relocation between 25 to 50 minutes of training per policy on average, improving over state-of-the-art results reported for similar tasks in the literature. These policies achieve perfect or near-perfect success rates, extreme robustness even under perturbations, and exhibit emergent recovery and correction behaviors. We hope that these promising results and our high-quality open-source implementation will provide a tool for the robotics community to facilitate further developments in robotic RL. Our code, documentation, and videos can be found at https://serl-robot.github.io/
In this paper, we scale evolutionary algorithms to high-dimensional optimization problems that deceptively possess a low effective dimensionality (certain dimensions do not significantly affect the objective function). To this end, an instantiation of the multiform optimization paradigm is presented, where multiple low-dimensional counterparts of a target high-dimensional task are generated via random embeddings. Since the exact relationship between the auxiliary (low-dimensional) tasks and the target is a priori unknown, a multiform evolutionary algorithm is developed for unifying all formulations into a single multi-task setting. The resultant joint optimization enables the target task to efficiently reuse solutions evolved across various low-dimensional searches via cross-form genetic transfers, hence speeding up overall convergence characteristics. To validate the overall efficacy of our proposed algorithmic framework, comprehensive experimental studies are carried out on well-known continuous benchmark functions as well as a set of practical problems in the hyper-parameter tuning of machine learning models and deep learning models in classification tasks and Predator-Prey games, respectively.
Transfer optimization enables data-efficient optimization of a target task by leveraging experiential priors from related source tasks. This is especially useful in multiobjective optimization settings where a set of trade-off solutions is sought under tight evaluation budgets. In this paper, we introduce a novel concept of inverse transfer in multiobjective optimization. Inverse transfer stands out by employing probabilistic inverse models to map performance vectors in the objective space to population search distributions in task-specific decision space, facilitating knowledge transfer through objective space unification. Building upon this idea, we introduce the first Inverse Transfer Multiobjective Evolutionary Optimizer (invTrEMO). A key highlight of invTrEMO is its ability to harness the common objective functions prevalent in many application areas, even when decision spaces do not precisely align between tasks. This allows invTrEMO to uniquely and effectively utilize information from heterogeneous source tasks as well. Furthermore, invTrEMO yields high-precision inverse models as a significant byproduct, enabling the generation of tailored solutions on-demand based on user preferences. Empirical studies on multi- and many-objective benchmark problems, as well as a practical case study, showcase the faster convergence rate and modelling accuracy of the invTrEMO relative to state-of-the-art evolutionary and Bayesian optimization algorithms. The source code of the invTrEMO is made available at https://github.com/LiuJ-2023/invTrEMO.
We study the problem of modeling a population of agents pursuing unknown goals subject to unknown computational constraints. In standard models of bounded rationality, sub-optimal decision-making is simulated by adding homoscedastic noise to optimal decisions rather than explicitly simulating constrained inference. In this work, we introduce a latent inference budget model (L-IBM) that models agents' computational constraints explicitly, via a latent variable (inferred jointly with a model of agents' goals) that controls the runtime of an iterative inference algorithm. L-IBMs make it possible to learn agent models using data from diverse populations of suboptimal actors. In three modeling tasks -- inferring navigation goals from routes, inferring communicative intents from human utterances, and predicting next moves in human chess games -- we show that L-IBMs match or outperform Boltzmann models of decision-making under uncertainty. Inferred inference budgets are themselves meaningful, efficient to compute, and correlated with measures of player skill, partner skill and task difficulty.
Physics-informed neural networks (PINNs) are at the forefront of scientific machine learning, making possible the creation of machine intelligence that is cognizant of physical laws and able to accurately simulate them. In this paper, the potential of discovering PINNs that generalize over an entire family of physics tasks is studied, for the first time, through a biological lens of the Baldwin effect. Drawing inspiration from the neurodevelopment of precocial species that have evolved to learn, predict and react quickly to their environment, we envision PINNs that are pre-wired with connection strengths inducing strong biases towards efficient learning of physics. To this end, evolutionary selection pressure (guided by proficiency over a family of tasks) is coupled with lifetime learning (to specialize on a smaller subset of those tasks) to produce PINNs that demonstrate fast and physics-compliant prediction capabilities across a range of empirically challenging problem instances. The Baldwinian approach achieves an order of magnitude improvement in prediction accuracy at a fraction of the computation cost compared to state-of-the-art results with PINNs meta-learned by gradient descent. This paper marks a leap forward in the meta-learning of PINNs as generalizable physics solvers.
With the increasingly widespread adoption of AI in healthcare, maintaining the accuracy and reliability of AI models in clinical practice has become crucial. In this context, we introduce novel methods for monitoring the performance of radiology AI classification models in practice, addressing the challenges of obtaining real-time ground truth for performance monitoring. We propose two metrics - predictive divergence and temporal stability - to be used for preemptive alerts of AI performance changes. Predictive divergence, measured using Kullback-Leibler and Jensen-Shannon divergences, evaluates model accuracy by comparing predictions with those of two supplementary models. Temporal stability is assessed through a comparison of current predictions against historical moving averages, identifying potential model decay or data drift. This approach was retrospectively validated using chest X-ray data from a single-center imaging clinic, demonstrating its effectiveness in maintaining AI model reliability. By providing continuous, real-time insights into model performance, our system ensures the safe and effective use of AI in clinical decision-making, paving the way for more robust AI integration in healthcare