Artem Zholus, Alexey Skrynnik, Shrestha Mohanty et al. · microsoft-research, mit
We present the IGLU Gridworld: a reinforcement learning environment for building and evaluating language conditioned embodied agents in a scalable way. The environment features visual agent embodiment, interactive learning through collaboration, language conditioned RL, and combinatorically hard task (3d blocks building) space.
Alexey Skrynnik, Anton Andreychuk, Konstantin Yakovlev et al.
We introduce POGEMA (https://github.com/AIRI-Institute/pogema) a sandbox for challenging partially observable multi-agent pathfinding (PO-MAPF) problems . This is a grid-based environment that was specifically designed to be a flexible, tunable and scalable benchmark. It can be tailored to a variety of PO-MAPF, which can serve as an excellent testing ground for planning and learning methods, and their combination, which will allow us to move towards filling the gap between AI planning and learning.
Dmitry Yudin, Yaroslav Solomentsev, Ruslan Musaev et al.
We present a novel dataset named as HPointLoc, specially designed for exploring capabilities of visual place recognition in indoor environment and loop detection in simultaneous localization and mapping. The loop detection sub-task is especially relevant when a robot with an on-board RGB-D camera can drive past the same place (``Point") at different angles. The dataset is based on the popular Habitat simulator, in which it is possible to generate photorealistic indoor scenes using both own sensor data and open datasets, such as Matterport3D. To study the main stages of solving the place recognition problem on the HPointLoc dataset, we proposed a new modular approach named as PNTR. It first performs an image retrieval with the Patch-NetVLAD method, then extracts keypoints and matches them using R2D2, LoFTR or SuperPoint with SuperGlue, and finally performs a camera pose optimization step with TEASER++. Such a solution to the place recognition problem has not been previously studied in existing publications. The PNTR approach has shown the best quality metrics on the HPointLoc dataset and has a high potential for real use in localization systems for unmanned vehicles. The proposed dataset and framework are publicly available: https://github.com/metra4ok/HPointLoc.
Alexandr V. Korchemnyi, Anatoly O. Onishchenko, Eva A. Bakaeva et al.
Embodied agents tasked with complex scenarios, whether in real or simulated environments, rely heavily on robust planning capabilities. When instructions are formulated in natural language, large language models (LLMs) equipped with extensive linguistic knowledge can play this role. However, to effectively exploit the ability of such models to handle linguistic ambiguity, to retrieve information from the environment, and to be based on the available skills of an agent, an appropriate architecture must be designed. We propose a Hierarchical Embodied Language Planner, called HELP, consisting of a set of LLM-based agents, each dedicated to solving a different subtask. We evaluate the proposed approach on a household task and perform real-world experiments with an embodied agent. We also focus on the use of open source LLMs with a relatively small number of parameters, to enable autonomous deployment.
Egor Cherepanov, Alexey Staroverov, Alexey K. Kovalev et al.
Transformers have become increasingly popular in offline reinforcement learning (RL) due to their ability to treat agent trajectories as sequences, reframing policy learning as a sequence modeling task. However, in partially observable environments (POMDPs), effective decision-making depends on retaining information about past events -- something that standard transformers struggle with due to the quadratic complexity of self-attention, which limits their context length. One solution to this problem is to extend transformers with memory mechanisms. We propose the Recurrent Action Transformer with Memory (RATE), a novel transformer-based architecture for offline RL that incorporates a recurrent memory mechanism designed to regulate information retention. We evaluate RATE across a diverse set of environments: memory-intensive tasks (ViZDoom-Two-Colors, T-Maze, Memory Maze, Minigrid-Memory, and POPGym), as well as standard Atari and MuJoCo benchmarks. Our comprehensive experiments demonstrate that RATE significantly improves performance in memory-dependent settings while remaining competitive on standard tasks across a broad range of baselines. These findings underscore the pivotal role of integrated memory mechanisms in offline RL and establish RATE as a unified, high-capacity architecture for effective decision-making over extended horizons. Code: https://sites.google.com/view/rate-model/.
Artem Latyshev, Aleksandr I. Panov
The reinforcement learning research area contains a wide range of methods for solving the problems of intelligent agent control. Despite the progress that has been made, the task of creating a highly autonomous agent is still a significant challenge. One potential solution to this problem is intrinsic motivation, a concept derived from developmental psychology. This review considers the existing methods for determining intrinsic motivation based on the world model obtained by the agent. We propose a systematic approach to current research in this field, which consists of three categories of methods, distinguished by the way they utilize a world model in the agent's components: complementary intrinsic reward, exploration policy, and intrinsically motivated goals. The proposed unified framework describes the architecture of agents using a world model and intrinsic motivation to improve learning. The potential for developing new techniques in this area of research is also examined.
Artem Tsypin, Leonid Ugadiarov, Kuzma Khrabrov et al.
Molecular conformation optimization is crucial to computer-aided drug discovery and materials design. Traditional energy minimization techniques rely on iterative optimization methods that use molecular forces calculated by a physical simulator (oracle) as anti-gradients. However, this is a computationally expensive approach that requires many interactions with a physical simulator. One way to accelerate this procedure is to replace the physical simulator with a neural network. Despite recent progress in neural networks for molecular conformation energy prediction, such models are prone to distribution shift, leading to inaccurate energy minimization. We find that the quality of energy minimization with neural networks can be improved by providing optimization trajectories as additional training data. Still, it takes around $5 \times 10^5$ additional conformations to match the physical simulator's optimization quality. In this work, we present the Gradual Optimization Learning Framework (GOLF) for energy minimization with neural networks that significantly reduces the required additional data. The framework consists of an efficient data-collecting scheme and an external optimizer. The external optimizer utilizes gradients from the energy prediction model to generate optimization trajectories, and the data-collecting scheme selects additional training data to be processed by the physical simulator. Our results demonstrate that the neural network trained with GOLF performs on par with the oracle on a benchmark of diverse drug-like molecules using $50$x less additional data.
Leonid Ugadiarov, Vitaliy Vorobyov, Aleksandr I. Panov
The advances in unsupervised object-centric representation learning have significantly improved its application to downstream tasks. Recent works highlight that disentangled object representations can aid policy learning in image-based, object-centric reinforcement learning tasks. This paper proposes a novel object-centric reinforcement learning algorithm that integrates actor-critic and model-based approaches by incorporating an object-centric world model within the critic. The world model captures the environment's data-generating process by predicting the next state and reward given the current state-action pair, where actions are interventions in the environment. In model-based reinforcement learning, world model learning can be interpreted as a causal induction problem, where the agent must learn the causal relationships underlying the environment's dynamics. We evaluate our method in a simulated 3D robotic environment and a 2D environment with compositional structure. As baselines, we compare against object-centric, model-free actor-critic algorithms and a state-of-the-art monolithic model-based algorithm. While the baselines show comparable performance in easier tasks, our approach outperforms them in more challenging scenarios with a large number of objects or more complex dynamics.
Zoya Volovikova, Alexey Skrynnik, Petr Kuderov et al.
In this study, we address the issue of enabling an artificial intelligence agent to execute complex language instructions within virtual environments. In our framework, we assume that these instructions involve intricate linguistic structures and multiple interdependent tasks that must be navigated successfully to achieve the desired outcomes. To effectively manage these complexities, we propose a hierarchical framework that combines the deep language comprehension of large language models with the adaptive action-execution capabilities of reinforcement learning agents. The language module (based on LLM) translates the language instruction into a high-level action plan, which is then executed by a pre-trained reinforcement learning agent. We have demonstrated the effectiveness of our approach in two different environments: in IGLU, where agents are instructed to build structures, and in Crafter, where agents perform tasks and interact with objects in the surrounding environment according to language commands.
Daniil Kirilenko, Vitaliy Vorobyov, Alexey K. Kovalev et al.
Object-centric architectures usually apply a differentiable module to the entire feature map to decompose it into sets of entity representations called slots. Some of these methods structurally resemble clustering algorithms, where the cluster's center in latent space serves as a slot representation. Slot Attention is an example of such a method, acting as a learnable analog of the soft k-means algorithm. Our work employs a learnable clustering method based on the Gaussian Mixture Model. Unlike other approaches, we represent slots not only as centers of clusters but also incorporate information about the distance between clusters and assigned vectors, leading to more expressive slot representations. Our experiments demonstrate that using this approach instead of Slot Attention improves performance in object-centric scenarios, achieving state-of-the-art results in the set property prediction task.
Evgenii Dzhivelikian, Petr Kuderov, Aleksandr I. Panov
This paper presents a novel approach to address the challenge of online sequence learning for decision making under uncertainty in non-stationary, partially observable environments. The proposed algorithm, Distributed Hebbian Temporal Memory (DHTM), is based on the factor graph formalism and a multi-component neuron model. DHTM aims to capture sequential data relationships and make cumulative predictions about future observations, forming Successor Features (SFs). Inspired by neurophysiological models of the neocortex, the algorithm uses distributed representations, sparse transition matrices, and local Hebbian-like learning rules to overcome the instability and slow learning of traditional temporal memory algorithms such as RNN and HMM. Experimental results show that DHTM outperforms LSTM, RWKV and a biologically inspired HMM-like algorithm, CSCG, on non-stationary data sets. Our results suggest that DHTM is a promising approach to address the challenges of online sequence learning and planning in dynamic environments.
Anatoly O. Onishchenko, Alexey K. Kovalev, Aleksandr I. Panov
Methods that use Large Language Models (LLM) as planners for embodied instruction following tasks have become widespread. To successfully complete tasks, the LLM must be grounded in the environment in which the robot operates. One solution is to use a scene graph that contains all the necessary information. Modern methods rely on prebuilt scene graphs and assume that all task-relevant information is available at the start of planning. However, these approaches do not account for changes in the environment that may occur between the graph construction and the task execution. We propose LookPlanGraph - a method that leverages a scene graph composed of static assets and object priors. During plan execution, LookPlanGraph continuously updates the graph with relevant objects, either by verifying existing priors or discovering new entities. This is achieved by processing the agents egocentric camera view using a Vision Language Model. We conducted experiments with changed object positions VirtualHome and OmniGibson simulated environments, demonstrating that LookPlanGraph outperforms methods based on predefined static scene graphs. To demonstrate the practical applicability of our approach, we also conducted experiments in a real-world setting. Additionally, we introduce the GraSIF (Graph Scenes for Instruction Following) dataset with automated validation framework, comprising 514 tasks drawn from SayPlan Office, BEHAVIOR-1K, and VirtualHome RobotHow. Project page available at https://lookplangraph.github.io .
Danil S. Grigorev, Alexey K. Kovalev, Aleksandr I. Panov
In the field of robotics, researchers face a critical challenge in ensuring reliable and efficient task planning. Verifying high-level task plans before execution significantly reduces errors and enhance the overall performance of these systems. In this paper, we propose an architecture for automatically verifying high-level task plans before their execution in simulator or real-world environments. Leveraging Large Language Models (LLMs), our approach consists of two key steps: first, the conversion of natural language instructions into Linear Temporal Logic (LTL), followed by a comprehensive analysis of action sequences. The module uses the reasoning capabilities of the LLM to evaluate logical coherence and identify potential gaps in the plan. Rigorous testing on datasets of varying complexity demonstrates the broad applicability of the module to household tasks. We contribute to improving the reliability and efficiency of task planning and addresses the critical need for robust pre-execution verification in autonomous systems. The code is available at https://verifyllm.github.io.
Oleg Shchendrigin, Egor Cherepanov, Alexey K. Kovalev et al.
Effective decision-making in the real world depends on memory that is both stable and adaptive: environments change over time, and agents must retain relevant information over long horizons while also updating or overwriting outdated content when circumstances shift. Existing Reinforcement Learning (RL) benchmarks and memory-augmented agents focus primarily on retention, leaving the equally critical ability of memory rewriting largely unexplored. To address this gap, we introduce a benchmark that explicitly tests continual memory updating under partial observability, i.e. the natural setting where an agent must rely on memory rather than current observations, and use it to compare recurrent, transformer-based, and structured memory architectures. Our experiments reveal that classic recurrent models, despite their simplicity, demonstrate greater flexibility and robustness in memory rewriting tasks than modern structured memories, which succeed only under narrow conditions, and transformer-based agents, which often fail beyond trivial retention cases. These findings expose a fundamental limitation of current approaches and emphasize the necessity of memory mechanisms that balance stable retention with adaptive updating. Our work highlights this overlooked challenge, introduces benchmarks to evaluate it, and offers insights for designing future RL agents with explicit and trainable forgetting mechanisms. Code: https://quartz-admirer.github.io/Memory-Rewriting/
Egor Cherepanov, Daniil Zelezetsky, Alexey K. Kovalev et al.
Pixel-based reinforcement learning agents often fail under purely visual distribution shift even when latent dynamics and rewards are unchanged, but existing benchmarks entangle multiple sources of shift and hinder systematic analysis. We introduce KAGE-Env, a JAX-native 2D platformer that factorizes the observation process into independently controllable visual axes while keeping the underlying control problem fixed. By construction, varying a visual axis affects performance only through the induced state-conditional action distribution of a pixel policy, providing a clean abstraction for visual generalization. Building on this environment, we define KAGE-Bench, a benchmark of six known-axis suites comprising 34 train-evaluation configuration pairs that isolate individual visual shifts. Using a standard PPO-CNN baseline, we observe strong axis-dependent failures, with background and photometric shifts often collapsing success, while agent-appearance shifts are comparatively benign. Several shifts preserve forward motion while breaking task completion, showing that return alone can obscure generalization failures. Finally, the fully vectorized JAX implementation enables up to 33M environment steps per second on a single GPU, enabling fast and reproducible sweeps over visual factors. Code: https://avanturist322.github.io/KAGEBench/.
Anastasiia Ivanova, Eva Bakaeva, Zoya Volovikova et al.
As a part of an embodied agent, Large Language Models (LLMs) are typically used for behavior planning given natural language instructions from the user. However, dealing with ambiguous instructions in real-world environments remains a challenge for LLMs. Various methods for task ambiguity detection have been proposed. However, it is difficult to compare them because they are tested on different datasets and there is no universal benchmark. For this reason, we propose AmbiK (Ambiguous Tasks in Kitchen Environment), the fully textual dataset of ambiguous instructions addressed to a robot in a kitchen environment. AmbiK was collected with the assistance of LLMs and is human-validated. It comprises 1000 pairs of ambiguous tasks and their unambiguous counterparts, categorized by ambiguity type (Human Preferences, Common Sense Knowledge, Safety), with environment descriptions, clarifying questions and answers, user intents, and task plans, for a total of 2000 tasks. We hope that AmbiK will enable researchers to perform a unified comparison of ambiguity detection methods. AmbiK is available at https://github.com/cog-model/AmbiK-dataset.
Leonid Ugadiarov, Alexey Skrynnik, Aleksandr I. Panov
Exploration is an essential part of reinforcement learning, which restricts the quality of learned policy. Hard-exploration environments are defined by huge state space and sparse rewards. In such conditions, an exhaustive exploration of the environment is often impossible, and the successful training of an agent requires a lot of interaction steps. In this paper, we propose an exploration method called Rollback-Explore (RbExplore), which utilizes the concept of the persistent Markov decision process, in which agents during training can roll back to visited states. We test our algorithm in the hard-exploration Prince of Persia game, without rewards and domain knowledge. At all used levels of the game, our agent outperforms or shows comparable results with state-of-the-art curiosity methods with knowledge-based intrinsic motivation: ICM and RND. An implementation of RbExplore can be found at https://github.com/cds-mipt/RbExplore.
Egor Cherepanov, Nikita Kachaev, Alexey K. Kovalev et al.
Memory is crucial for enabling agents to tackle complex tasks with temporal and spatial dependencies. While many reinforcement learning (RL) algorithms incorporate memory, the field lacks a universal benchmark to assess an agent's memory capabilities across diverse scenarios. This gap is particularly evident in tabletop robotic manipulation, where memory is essential for solving tasks with partial observability and ensuring robust performance, yet no standardized benchmarks exist. To address this, we introduce MIKASA (Memory-Intensive Skills Assessment Suite for Agents), a comprehensive benchmark for memory RL, with three key contributions: (1) we propose a comprehensive classification framework for memory-intensive RL tasks, (2) we collect MIKASA-Base -- a unified benchmark that enables systematic evaluation of memory-enhanced agents across diverse scenarios, and (3) we develop MIKASA-Robo (pip install mikasa-robo-suite) -- a novel benchmark of 32 carefully designed memory-intensive tasks that assess memory capabilities in tabletop robotic manipulation. Our work introduces a unified framework to advance memory RL research, enabling more robust systems for real-world use. MIKASA is available at https://tinyurl.com/membenchrobots.
Maxim A. Patratskiy, Alexey K. Kovalev, Aleksandr I. Panov
Vision-Language-Action models have demonstrated remarkable capabilities in predicting agent movements within virtual environments and real-world scenarios based on visual observations and textual instructions. Although recent research has focused on enhancing spatial and temporal understanding independently, this paper presents a novel approach that integrates both aspects through visual prompting. We introduce a method that projects visual traces of key points from observations onto depth maps, enabling models to capture both spatial and temporal information simultaneously. The experiments in SimplerEnv show that the mean number of tasks successfully solved increased for 4% compared to SpatialVLA and 19% compared to TraceVLA. Furthermore, we show that this enhancement can be achieved with minimal training data, making it particularly valuable for real-world applications where data collection is challenging. The project page is available at https://ampiromax.github.io/ST-VLA.
Egor Cherepanov, Nikita Kachaev, Artem Zholus et al.
The incorporation of memory into agents is essential for numerous tasks within the domain of Reinforcement Learning (RL). In particular, memory is paramount for tasks that require the utilization of past information, adaptation to novel environments, and improved sample efficiency. However, the term ``memory'' encompasses a wide range of concepts, which, coupled with the lack of a unified methodology for validating an agent's memory, leads to erroneous judgments about agents' memory capabilities and prevents objective comparison with other memory-enhanced agents. This paper aims to streamline the concept of memory in RL by providing practical precise definitions of agent memory types, such as long-term versus short-term memory and declarative versus procedural memory, inspired by cognitive science. Using these definitions, we categorize different classes of agent memory, propose a robust experimental methodology for evaluating the memory capabilities of RL agents, and standardize evaluations. Furthermore, we empirically demonstrate the importance of adhering to the proposed methodology when evaluating different types of agent memory by conducting experiments with different RL agents and what its violation leads to.
Nikita Kachaev, Daniil Zelezetsky, Egor Cherepanov et al.
Despite their effectiveness and popularity in offline or model-based reinforcement learning (RL), transformers remain underexplored in online model-free RL due to their sensitivity to training setups and model design decisions such as how to structure the policy and value networks, share components, or handle temporal information. In this paper, we show that transformers can be strong baselines for continuous control in online model-free RL. We investigate key design questions: how to condition inputs, share components between actor and critic, and slice sequential data for training. Our experiments reveal stable architectural and training strategies enabling competitive performance across fully and partially observable tasks, and in both vector- and image-based settings. These findings offer practical guidance for applying transformers in online RL.
Nikita Kachaev, Mikhail Kolosov, Daniil Zelezetsky et al.
The growing success of Vision-Language-Action (VLA) models stems from the promise that pretrained Vision-Language Models (VLMs) can endow agents with transferable world knowledge and vision-language (VL) grounding, laying a foundation for action models with broader generalization. Yet when these VLMs are adapted to the action modality, it remains unclear to what extent their original VL representations and knowledge are preserved. In this work, we conduct a systematic study of representation retention during VLA fine-tuning, showing that naive action fine-tuning leads to degradation of visual representations. To characterize and measure these effects, we probe VLA's hidden representations and analyze attention maps, further, we design a set of targeted tasks and methods that contrast VLA models with their counterpart VLMs, isolating changes in VL capabilities induced by action fine-tuning. We further evaluate a range of strategies for aligning visual representations and introduce a simple yet effective method that mitigates degradation and yields improved generalization to out-of-distribution (OOD) scenarios. Taken together, our analysis clarifies the trade-off between action fine-tuning and the degradation of VL representations and highlights practical approaches to recover inherited VL capabilities. Code is publicly available: https://blind-vla-paper.github.io
Daniil Zelezetsky, Egor Cherepanov, Alexey K. Kovalev et al.
Offline reinforcement learning (RL) often deals with suboptimal data when collecting large expert datasets is unavailable or impractical. This limitation makes it difficult for agents to generalize and achieve high performance, as they must learn primarily from imperfect or inconsistent trajectories. A central challenge is therefore how to best leverage scarce expert demonstrations alongside abundant but lower-quality data. We demonstrate that incorporating even a tiny amount of expert experience can substantially improve RL agent performance. We introduce Re:Frame (Retrieving Experience From Associative Memory), a plug-in module that augments a standard offline RL policy (e.g., Decision Transformer) with a small external Associative Memory Buffer (AMB) populated by expert trajectories drawn from a separate dataset. During training on low-quality data, the policy learns to retrieve expert data from the Associative Memory Buffer (AMB) via content-based associations and integrate them into decision-making; the same AMB is queried at evaluation. This requires no environment interaction and no modifications to the backbone architecture. On D4RL MuJoCo tasks, using as few as 60 expert trajectories (0.1% of a 6000-trajectory dataset), Re:Frame consistently improves over a strong Decision Transformer baseline in three of four settings, with gains up to +10.7 normalized points. These results show that Re:Frame offers a simple and data-efficient way to inject scarce expert knowledge and substantially improve offline RL from low-quality datasets.
Zoya Volovikova, Gregory Gorbov, Petr Kuderov et al.
Following instructions in real-world conditions requires the ability to adapt to the world's volatility and entanglement: the environment is dynamic and unpredictable, instructions can be linguistically complex with diverse vocabulary, and the number of possible goals an agent may encounter is vast. Despite extensive research in this area, most studies are conducted in static environments with simple instructions and a limited vocabulary, making it difficult to assess agent performance in more diverse and challenging settings. To address this gap, we introduce CrafText, a benchmark for evaluating instruction following in a multimodal environment with diverse instructions and dynamic interactions. CrafText includes 3,924 instructions with 3,423 unique words, covering Localization, Conditional, Building, and Achievement tasks. Additionally, we propose an evaluation protocol that measures an agent's ability to generalize to novel instruction formulations and dynamically evolving task configurations, providing a rigorous test of both linguistic understanding and adaptive decision-making.
Alexandr Korchemnyi, Alexey K. Kovalev, Aleksandr I. Panov
The idea of disentangled representations is to reduce the data to a set of generative factors that produce it. Typically, such representations are vectors in latent space, where each coordinate corresponds to one of the generative factors. The object can then be modified by changing the value of a particular coordinate, but it is necessary to determine which coordinate corresponds to the desired generative factor -- a difficult task if the vector representation has a high dimension. In this article, we propose ArSyD (Architecture for Symbolic Disentanglement), which represents each generative factor as a vector of the same dimension as the resulting representation. In ArSyD, the object representation is obtained as a superposition of the generative factor vector representations. We call such a representation a \textit{symbolic disentangled representation}. We use the principles of Hyperdimensional Computing (also known as Vector Symbolic Architectures), where symbols are represented as hypervectors, allowing vector operations on them. Disentanglement is achieved by construction, no additional assumptions about the underlying distributions are made during training, and the model is only trained to reconstruct images in a weakly supervised manner. We study ArSyD on the dSprites and CLEVR datasets and provide a comprehensive analysis of the learned symbolic disentangled representations. We also propose new disentanglement metrics that allow comparison of methods using latent representations of different dimensions. ArSyD allows to edit the object properties in a controlled and interpretable way, and the dimensionality of the object property representation coincides with the dimensionality of the object representation itself.
Egor Cherepanov, Alexey K. Kovalev, Aleksandr I. Panov
Real-world robotic agents must act under partial observability and long horizons, where key cues may appear long before they affect decision making. However, most modern approaches rely solely on instantaneous information, without incorporating insights from the past. Standard recurrent or transformer models struggle with retaining and leveraging long-term dependencies: context windows truncate history, while naive memory extensions fail under scale and sparsity. We propose ELMUR (External Layer Memory with Update/Rewrite), a transformer architecture with structured external memory. Each layer maintains memory embeddings, interacts with them via bidirectional cross-attention, and updates them through an Least Recently Used (LRU) memory module using replacement or convex blending. ELMUR extends effective horizons up to 100,000 times beyond the attention window and achieves a 100% success rate on a synthetic T-Maze task with corridors up to one million steps. In POPGym, it outperforms baselines on more than half of the tasks. On MIKASA-Robo sparse-reward manipulation tasks with visual observations, it nearly doubles the performance of strong baselines. These results demonstrate that structured, layer-local external memory offers a simple and scalable approach to decision making under partial observability.
Daniil Zelezetsky, Alexey K. Kovalev, Aleksandr I. Panov
The appearance of transformer-based models in Reinforcement Learning (RL) has expanded the horizons of possibilities in robotics tasks, but it has simultaneously brought a wide range of challenges during its implementation, especially in model-free online RL. Some of the existing learning algorithms cannot be easily implemented with transformer-based models due to the instability of the latter. In this paper, we propose a method that uses the Accelerator policy as a transformer's trainer. The Accelerator, a simpler and more stable model, interacts with the environment independently while simultaneously training the transformer through behavior cloning during the first stage of the proposed algorithm. In the second stage, the pretrained transformer starts to interact with the environment in a fully online setting. As a result, this model-free algorithm accelerates the transformer in terms of its performance and helps it to train online in a more stable and faster way. By conducting experiments on both state-based and image-based ManiSkill environments, as well as on MuJoCo tasks in MDP and POMDP settings, we show that applying our algorithm not only enables stable training of transformers but also reduces training time on image-based environments by up to a factor of two. Moreover, it decreases the required replay buffer size in off-policy methods to 10-20 thousand, which significantly lowers the overall computational demands.
Nikita Kachaev, Andrei Spiridonov, Andrey Gorodetsky et al.
Benchmarks are crucial for evaluating progress in robotics and embodied AI. However, a significant gap exists between benchmarks designed for high-level language instruction following, which often assume perfect low-level execution, and those for low-level robot control, which rely on simple, one-step commands. This disconnect prevents a comprehensive evaluation of integrated systems where both task planning and physical execution are critical. To address this, we propose Kitchen-R, a novel benchmark that unifies the evaluation of task planning and low-level control within a simulated kitchen environment. Built as a digital twin using the Isaac Sim simulator and featuring more than 500 complex language instructions, Kitchen-R supports a mobile manipulator robot. We provide baseline methods for our benchmark, including a task-planning strategy based on a vision-language model and a low-level control policy based on diffusion policy. We also provide a trajectory collection system. Our benchmark offers a flexible framework for three evaluation modes: independent assessment of the planning module, independent assessment of the control policy, and, crucially, an integrated evaluation of the whole system. Kitchen-R bridges a key gap in embodied AI research, enabling more holistic and realistic benchmarking of language-guided robotic agents.
Artem Latyshev, Gregory Gorbov, Aleksandr I. Panov
Reinforcement Learning (RL) applications in real-world scenarios must prioritize safety and reliability, which impose strict constraints on agent behavior. Model-based RL leverages predictive world models for action planning and policy optimization, but inherent model inaccuracies can lead to catastrophic failures in safety-critical settings. We propose a novel model-based RL framework that jointly optimizes task performance and safety. To address world model errors, our method incorporates an adaptive mechanism that dynamically switches between model-based planning and direct policy execution. We resolve the objective mismatch problem of traditional model-based approaches using an implicit world model. Furthermore, our framework employs dynamic safety thresholds that adapt to the agent's evolving capabilities, consistently selecting actions that surpass safe policy suggestions in both performance and safety. Experiments demonstrate significant improvements over non-adaptive methods, showing that our approach optimizes safety and performance simultaneously rather than merely meeting minimum safety requirements. The proposed framework achieves robust performance on diverse safety-critical continuous control tasks, outperforming existing methods.
Artem Zholus, Aleksandr I. Panov
Model-based reinforcement learning (MBRL) allows solving complex tasks in a sample-efficient manner. However, no information is reused between the tasks. In this work, we propose a meta-learned addressing model called RAMa that provides training samples for the MBRL agent taken from continuously growing task-agnostic storage. The model is trained to maximize the expected agent's performance by selecting promising trajectories solving prior tasks from the storage. We show that such retrospective exploration can accelerate the learning process of the MBRL agent by better informing learned dynamics and prompting agent with exploratory trajectories. We test the performance of our approach on several domains from the DeepMind control suite, from Metaworld multitask benchmark, and from our bespoke environment implemented with a robotic NVIDIA Isaac simulator to test the ability of the model to act in a photorealistic, ray-traced environment.
Aleksey Staroverov, Aleksandr I. Panov
This work studies object goal navigation task, which involves navigating to the closest object related to the given semantic category in unseen environments. Recent works have shown significant achievements both in the end-to-end Reinforcement Learning approach and modular systems, but need a big step forward to be robust and optimal. We propose a hierarchical method that incorporates standard task formulation and additional area knowledge as landmarks, with a way to extract these landmarks. In a hierarchy, a low level consists of separately trained algorithms to the most intuitive skills, and a high level decides which skill is needed at this moment. With all proposed solutions, we achieve a 0.75 success rate in a realistic Habitat simulator. After a small stage of additional model training in a reconstructed virtual area at a simulator, we successfully confirmed our results in a real-world case.
Vasilii Davydov, Alexey Skrynnik, Konstantin Yakovlev et al.
In this paper, we consider the problem of multi-agent navigation in partially observable grid environments. This problem is challenging for centralized planning approaches as they, typically, rely on the full knowledge of the environment. We suggest utilizing the reinforcement learning approach when the agents, first, learn the policies that map observations to actions and then follow these policies to reach their goals. To tackle the challenge associated with learning cooperative behavior, i.e. in many cases agents need to yield to each other to accomplish a mission, we use a mixing Q-network that complements learning individual policies. In the experimental evaluation, we show that such approach leads to plausible results and scales well to large number of agents.
Andrey Gorodetskiy, Alexandra Shlychkova, Aleksandr I. Panov
This work is devoted to unresolved problems of Artificial General Intelligence - the inefficiency of transfer learning. One of the mechanisms that are used to solve this problem in the area of reinforcement learning is a model-based approach. In the paper we are expanding the schema networks method which allows to extract the logical relationships between objects and actions from the environment data. We present algorithms for training a Delta Schema Network (DSN), predicting future states of the environment and planning actions that will lead to positive reward. DSN shows strong performance of transfer learning on the classic Atari game environment.
Alexey Skrynnik, Aleksey Staroverov, Ermek Aitygulov et al.
Currently, deep reinforcement learning (RL) shows impressive results in complex gaming and robotic environments. Often these results are achieved at the expense of huge computational costs and require an incredible number of episodes of interaction between the agent and the environment. There are two main approaches to improving the sample efficiency of reinforcement learning methods - using hierarchical methods and expert demonstrations. In this paper, we propose a combination of these approaches that allow the agent to use low-quality demonstrations in complex vision-based environments with multiple related goals. Our forgetful experience replay (ForgER) algorithm effectively handles errors in expert data and reduces quality losses when adapting the action space and states representation to the agent's capabilities. Our proposed goal-oriented structuring of replay buffer allows the agent to automatically highlight sub-goals for solving complex hierarchical tasks in demonstrations. Our method is universal and can be integrated into various off-policy methods. It surpasses all known existing state-of-the-art RL methods using expert demonstrations on various model environments. The solution based on our algorithm beats all the solutions for the famous MineRL competition and allows the agent to mine a diamond in the Minecraft environment.
Alexey Skrynnik, Aleksey Staroverov, Ermek Aitygulov et al.
We present Hierarchical Deep Q-Network (HDQfD) that took first place in the MineRL competition. HDQfD works on imperfect demonstrations and utilizes the hierarchical structure of expert trajectories. We introduce the procedure of extracting an effective sequence of meta-actions and subgoals from demonstration data. We present a structured task-dependent replay buffer and adaptive prioritizing technique that allow the HDQfD agent to gradually erase poor-quality expert data from the buffer. In this paper, we present the details of the HDQfD algorithm and give the experimental results in the Minecraft domain.
Aleksandr I. Panov, Aleksey Skrynnik
We introduce a new approach to hierarchy formation and task decomposition in hierarchical reinforcement learning. Our method is based on the Hierarchy Of Abstract Machines (HAM) framework because HAM approach is able to design efficient controllers that will realize specific behaviors in real robots. The key to our algorithm is the introduction of the internal or "mental" environment in which the state represents the structure of the HAM hierarchy. The internal action in this environment leads to changes the hierarchy of HAMs. We propose the classical Q-learning procedure in the internal environment which allows the agent to obtain an optimal hierarchy. We extends the HAM framework by adding on-model approach to select the appropriate sub-machine to execute action sequences for certain class of external environment states. Preliminary experiments demonstrated the prospects of the method.
Aleksandr I. Panov, Konstantin Yakovlev
Behavior planning is known to be one of the basic cognitive functions, which is essential for any cognitive architecture of any control system used in robotics. At the same time most of the widespread planning algorithms employed in those systems are developed using only approaches and models of Artificial Intelligence and don't take into account numerous results of cognitive experiments. As a result, there is a strong need for novel methods of behavior planning suitable for modern cognitive architectures aimed at robot control. One such method is presented in this work and is studied within a special class of navigation task called smart relocation task. The method is based on the hierarchical two-level model of abstraction and knowledge representation, e.g. symbolic and subsymbolic. On the symbolic level sign world model is used for knowledge representation and hierarchical planning algorithm, PMA, is utilized for planning. On the subsymbolic level the task of path planning is considered and solved as a graph search problem. Interaction between both planners is examined and inter-level interfaces and feedback loops are described. Preliminary experimental results are presented.
Aleksandr I. Panov, Konstantin Yakovlev
In this paper we outline the approach of solving special type of navigation tasks for robotic systems, when a coalition of robots (agents) acts in the 2D environment, which can be modified by the actions, and share the same goal location. The latter is originally unreachable for some members of the coalition, but the common task still can be accomplished as the agents can assist each other (e.g. by modifying the environment). We call such tasks smart relocation tasks (as the can not be solved by pure path planning methods) and study spatial and behavior interaction of robots while solving them. We use cognitive approach and introduce semiotic knowledge representation - sign world model which underlines behavioral planning methodology. Planning is viewed as a recursive search process in the hierarchical state-space induced by sings with path planning signs reside on the lowest level. Reaching this level triggers path planning which is accomplished by state of the art grid-based planners focused on producing smooth paths (e.g. LIAN) and thus indirectly guarantying feasibility of that paths against agent's dynamic constraints.