Marek Cygan

Piotr Krzywicki, Krzysztof Ciebiera, Rafał Michaluk et al.

Gathering real-world data from the robot quickly becomes a bottleneck when constructing a robot learning system for grasping. In this work, we design a semi-supervised grasping system that, on top of a small sample of robot experience, takes advantage of images of products to be picked, which are collected without any interactions with the robot. We validate our findings both in the simulation and in the real world. In the regime of a small number of robot training samples, taking advantage of the unlabeled data allows us to achieve performance at the level of 10-fold bigger dataset size used by the baseline. The code and datasets used in the paper will be released at https://github.com/nomagiclab/grasping-student.

Dominik Filipiak, Andrzej Zapała, Piotr Tempczyk et al.

We present Polite Teacher, a simple yet effective method for the task of semi-supervised instance segmentation. The proposed architecture relies on the Teacher-Student mutual learning framework. To filter out noisy pseudo-labels, we use confidence thresholding for bounding boxes and mask scoring for masks. The approach has been tested with CenterMask, a single-stage anchor-free detector. Tested on the COCO 2017 val dataset, our architecture significantly (approx. +8 pp. in mask AP) outperforms the baseline at different supervision regimes. To the best of our knowledge, this is one of the first works tackling the problem of semi-supervised instance segmentation and the first one devoted to an anchor-free detector.

Michal Nauman, Marek Cygan

We study the variance of stochastic policy gradients (SPGs) with many action samples per state. We derive a many-actions optimality condition, which determines when many-actions SPG yields lower variance as compared to a single-action agent with proportionally extended trajectory. We propose Model-Based Many-Actions (MBMA), an approach leveraging dynamics models for many-actions sampling in the context of SPG. MBMA addresses issues associated with existing implementations of many-actions SPG and yields lower bias and comparable variance to SPG estimated from states in model-simulated rollouts. We find that MBMA bias and variance structure matches that predicted by theory. As a result, MBMA achieves improved sample efficiency and higher returns on a range of continuous action environments as compared to model-free, many-actions, and model-based on-policy SPG baselines.

Piotr Tempczyk, Ksawery Smoczyński, Philip Smolenski-Jensen et al.

One of the most popular estimation methods in Bayesian neural networks (BNN) is mean-field variational inference (MFVI). In this work, we show that neural networks with ReLU activation function induce posteriors, that are hard to fit with MFVI. We provide a theoretical justification for this phenomenon, study it empirically, and report the results of a series of experiments to investigate the effect of activation function on the calibration of BNNs. We find that using Leaky ReLU activations leads to more Gaussian-like weight posteriors and achieves a lower expected calibration error (ECE) than its ReLU-based counterpart.

Kevin Qiu, Kyle Walker, Mike Y. Michelis et al.

We present Swim2Real, a pipeline that calibrates a 16-parameter robotic fish simulator from swimming videos using vision-language model (VLM) feedback, requiring no hand-designed search stages. Calibrating soft aquatic robots is particularly challenging because nonlinear fluid-structure coupling makes the parameter landscape chaotic, simplified fluid models introduce a persistent sim-to-real gap, and controlled aquatic experiments are difficult to reproduce. Prior work on this platform required three manually tailored stages to handle this complexity. The VLM compares simulated and real videos and proposes parameter updates. A backtracking line search then validates each step size, tripling the accept rate from 14% to 42% by recovering proposals where the direction is correct but the magnitude is too large. Swim2Real calibrates all 16 parameters simultaneously, most closely matching real fish velocities across all motor frequencies (MAE = 7.4 mm/s, 43% lower than the next-best method), with zero outlier seeds across five runs. Motor commands from the trained policy transfer to the physical fish at 50 Hz, completing the pipeline from swimming video to real-world deployment. Downstream RL policies swim 12% farther than those from BayesOpt-calibrated simulators and 90% farther than CMA-ES. These results demonstrate that VLM-guided calibration can close the sim-to-real gap for aquatic robots directly from video, enabling zero-shot RL transfer to physical swimmers without manual system identification, a step toward automated, general-purpose simulator tuning for underwater robotics.

Szymon Antoniak, Michał Krutul, Maciej Pióro et al.

Mixture of Experts (MoE) models based on Transformer architecture are pushing the boundaries of language and vision tasks. The allure of these models lies in their ability to substantially increase the parameter count without a corresponding increase in FLOPs. Most widely adopted MoE models are discontinuous with respect to their parameters - often referred to as sparse. At the same time, existing continuous MoE designs either lag behind their sparse counterparts or are incompatible with autoregressive decoding. Motivated by the observation that the adaptation of fully continuous methods has been an overarching trend in deep learning, we develop Mixture of Tokens (MoT), a simple, continuous architecture that is capable of scaling the number of parameters similarly to sparse MoE models. Unlike conventional methods, MoT assigns mixtures of tokens from different examples to each expert. This architecture is fully compatible with autoregressive training and generation. Our best models not only achieve a 3x increase in training speed over dense Transformer models in language pretraining but also match the performance of state-of-the-art MoE architectures. Additionally, a close connection between MoT and MoE is demonstrated through a novel technique we call transition tuning.

Kevin Qiu, Władysław Pałucki, Krzysztof Ciebiera et al.

We introduce RoboMorph, an automated approach for generating and optimizing modular robot designs using large language models (LLMs) and evolutionary algorithms. In this framework, we represent each robot design as a grammar and leverage the capabilities of LLMs to navigate the extensive robot design space, which is traditionally time-consuming and computationally demanding. By introducing a best-shot prompting technique and a reinforcement learning-based control algorithm, RoboMorph iteratively improves robot designs through feedback loops. Experimental results demonstrate that RoboMorph successfully generates nontrivial robots optimized for different terrains while showcasing improvements in robot morphology over successive evolutions. Our approach highlights the potential of using LLMs for data-driven, modular robot design, providing a promising methodology that can be extended to other domains with similar design frameworks.

Michal Nauman, Marek Cygan

Risk-aware Reinforcement Learning (RL) algorithms like SAC and TD3 were shown empirically to outperform their risk-neutral counterparts in a variety of continuous-action tasks. However, the theoretical basis for the pessimistic objectives these algorithms employ remains unestablished, raising questions about the specific class of policies they are implementing. In this work, we apply the expected utility hypothesis, a fundamental concept in economics, to illustrate that both risk-neutral and risk-aware RL goals can be interpreted through expected utility maximization using an exponential utility function. This approach reveals that risk-aware policies effectively maximize value certainty equivalent, aligning them with conventional decision theory principles. Furthermore, we propose Dual Actor-Critic (DAC). DAC is a risk-aware, model-free algorithm that features two distinct actor networks: a pessimistic actor for temporal-difference learning and an optimistic actor for exploration. Our evaluations of DAC across various locomotion and manipulation tasks demonstrate improvements in sample efficiency and final performance. Remarkably, DAC, while requiring significantly less computational resources, matches the performance of leading model-based methods in the complex dog and humanoid domains.

Kevin Qiu, Yu Zhang, Marek Cygan et al.

Calibrating a robot simulator's physics parameters (friction, damping, material stiffness) to match real hardware is often done by hand or with black-box optimizers that reduce error but cannot explain which physical discrepancies drive the error. When sensing is limited to external cameras, the problem is further compounded by perception noise and the absence of direct force or state measurements. We present Vid2Sid, a video-driven system identification pipeline that couples foundation-model perception with a VLM-in-the-loop optimizer that analyzes paired sim-real videos, diagnoses concrete mismatches, and proposes physics parameter updates with natural language rationales. We evaluate our approach on a tendon-actuated finger (rigid-body dynamics in MuJoCo) and a deformable continuum tentacle (soft-body dynamics in PyElastica). On sim2real holdout controls unseen during training, Vid2Sid achieves the best average rank across all settings, matching or exceeding black-box optimizers while uniquely providing interpretable reasoning at each iteration. Sim2sim validation confirms that Vid2Sid recovers ground-truth parameters most accurately (mean relative error under 13\% vs. 28--98\%), and ablation analysis reveals three calibration regimes. VLM-guided optimization excels when perception is clean and the simulator is expressive, while model-class limitations bound performance in more challenging settings.

Jakub Krajewski, Jan Ludziejewski, Kamil Adamczewski et al.

Mixture of Experts (MoE) models have emerged as a primary solution for reducing the computational cost of Large Language Models. In this work, we analyze their scaling properties, incorporating an expanded range of variables. Specifically, we introduce a new hyperparameter, granularity, whose adjustment enables precise control over the size of the experts. Building on this, we establish scaling laws for fine-grained MoE, taking into account the number of training tokens, model size, and granularity. Leveraging these laws, we derive the optimal training configuration for a given computational budget. Our findings not only show that MoE models consistently outperform dense Transformers but also highlight that the efficiency gap between dense and MoE models widens as we scale up the model size and training budget. Furthermore, we demonstrate that the common practice of setting the size of experts in MoE to mirror the feed-forward layer is not optimal at almost any computational budget.

Jan Małaśnicki, Kamil Ciebiera, Mateusz Boruń et al.

Recent years have seen a growing interest and adoption of LLMs, with Mixture-of-Experts (MoE) emerging as a leading architecture in extremely large models. Currently, the largest open-source models reach over $1$T parameters. At such scales, hyperparameter tuning becomes prohibitively expensive. Precisely for this reason, the $μ$Transfer is becoming a key technique. It allows for seamless transfer of optimal hyperparameters across model scales, resulting in a huge reduction in tuning costs. However, existing work has primarily focused on dense LLMs, leaving MoE architectures unexplored. In this work, we derive a $μ$-Parameterization for MoE, providing theoretical guarantees for feature learning across model widths. Our experiments demonstrate that the optimal learning rate reliably transfers across model sizes, establishing a foundation for efficient hyperparameter tuning in large-scale MoE models.

Maciej Pióro, Kamil Ciebiera, Krystian Król et al.

State Space Models (SSMs) have become serious contenders in the field of sequential modeling, challenging the dominance of Transformers. At the same time, Mixture of Experts (MoE) has significantly improved Transformer-based Large Language Models, including recent state-of-the-art open models. We propose that to unlock the potential of SSMs for scaling, they should be combined with MoE. We showcase this on Mamba, a recent SSM-based model that achieves remarkable performance. Our model, MoE-Mamba, outperforms both Mamba and baseline Transformer-MoE. In particular, MoE-Mamba reaches the same performance as Mamba in $2.35\times$ fewer training steps while preserving the inference performance gains of Mamba against Transformer.

Michal Nauman, Michał Bortkiewicz, Piotr Miłoś et al.

Recent advancements in off-policy Reinforcement Learning (RL) have significantly improved sample efficiency, primarily due to the incorporation of various forms of regularization that enable more gradient update steps than traditional agents. However, many of these techniques have been tested in limited settings, often on tasks from single simulation benchmarks and against well-known algorithms rather than a range of regularization approaches. This limits our understanding of the specific mechanisms driving RL improvements. To address this, we implemented over 60 different off-policy agents, each integrating established regularization techniques from recent state-of-the-art algorithms. We tested these agents across 14 diverse tasks from 2 simulation benchmarks, measuring training metrics related to overestimation, overfitting, and plasticity loss -- issues that motivate the examined regularization techniques. Our findings reveal that while the effectiveness of a specific regularization setup varies with the task, certain combinations consistently demonstrate robust and superior performance. Notably, a simple Soft Actor-Critic agent, appropriately regularized, reliably finds a better-performing policy within the training regime, which previously was achieved mainly through model-based approaches.

Jan Ludziejewski, Maciej Pióro, Jakub Krajewski et al.

Mixture of Experts (MoE) architectures have significantly increased computational efficiency in both research and real-world applications of large-scale machine learning models. However, their scalability and efficiency under memory constraints remain relatively underexplored. In this work, we present joint scaling laws for dense and MoE models, incorporating key factors such as the number of active parameters, dataset size, and the number of experts. Our findings provide a principled framework for selecting the optimal MoE configuration under fixed memory and compute budgets. Surprisingly, we show that MoE models can be more memory-efficient than dense models, contradicting conventional wisdom. To derive and validate the theoretical predictions of our scaling laws, we conduct over 280 experiments with up to 2.7B active parameters and up to 5B total parameters. These results offer actionable insights for designing and deploying MoE models in practical large-scale training scenarios.

Michal Nauman, Marek Cygan, Carmelo Sferrazza et al.

Recent advances in language modeling and vision stem from training large models on diverse, multi-task data. This paradigm has had limited impact in value-based reinforcement learning (RL), where improvements are often driven by small models trained in a single-task context. This is because in multi-task RL sparse rewards and gradient conflicts make optimization of temporal difference brittle. Practical workflows for generalist policies therefore avoid online training, instead cloning expert trajectories or distilling collections of single-task policies into one agent. In this work, we show that the use of high-capacity value models trained via cross-entropy and conditioned on learnable task embeddings addresses the problem of task interference in online RL, allowing for robust and scalable multi-task training. We test our approach on 7 multi-task benchmarks with over 280 unique tasks, spanning high degree-of-freedom humanoid control and discrete vision-based RL. We find that, despite its simplicity, the proposed approach leads to state-of-the-art single and multi-task performance, as well as sample-efficient transfer to new tasks.

Adam Pardyl, Dominik Matuszek, Mateusz Przebieracz et al.

The real world is messy and unstructured. Uncovering critical information often requires active, goal-driven exploration. It remains to be seen whether Vision-Language Models (VLMs), which recently emerged as a popular zero-shot tool in many difficult tasks, can operate effectively in such conditions. In this paper, we answer this question by introducing FlySearch, a 3D, outdoor, photorealistic environment for searching and navigating to objects in complex scenes. We define three sets of scenarios with varying difficulty and observe that state-of-the-art VLMs cannot reliably solve even the simplest exploration tasks, with the gap to human performance increasing as the tasks get harder. We identify a set of central causes, ranging from vision hallucination, through context misunderstanding, to task planning failures, and we show that some of them can be addressed by finetuning. We publicly release the benchmark, scenarios, and the underlying codebase.

Michal Nauman, Marek Cygan, Pieter Abbeel

RL agents are typically trained under a single, fixed reward function, which makes them brittle to reward misspecification and limits their ability to adapt to changing task preferences. We introduce Reward-Conditioned Reinforcement Learning (RCRL), a framework that trains a single agent to optimize a family of reward specifications while collecting experience under only one nominal objective. RCRL conditions the agent on reward parameterizations and learns multiple reward objectives from a shared replay data entirely off-policy, enabling a single policy to represent reward-specific behaviors. Across single-task, multi-task, and vision-based benchmarks, we show that RCRL not only improves performance under the nominal reward parameterization, but also enables efficient adaptation to new parameterizations. Our results demonstrate that RCRL provides a scalable mechanism for learning robust, steerable policies without sacrificing the simplicity of single-task training.

Kevin Qiu, Marek Cygan

Automating the co-design of a robot's morphology and control is a long-standing challenge due to the vast design space and the tight coupling between body and behavior. We introduce Debate2Create (D2C), a framework in which large language model (LLM) agents engage in a structured dialectical debate to jointly optimize a robot's design and its reward function. In each round, a design agent proposes targeted morphological modifications, and a control agent devises a reward function tailored to exploit the new design. A panel of pluralistic judges then evaluates the design-control pair in simulation and provides feedback that guides the next round of debate. Through iterative debates, the agents progressively refine their proposals, producing increasingly effective robot designs. Notably, D2C yields diverse and specialized morphologies despite no explicit diversity objective. On a quadruped locomotion benchmark, D2C discovers designs that travel 73% farther than the default, demonstrating that structured LLM-based debate can serve as a powerful mechanism for emergent robot co-design. Our results suggest that multi-agent debate, when coupled with physics-grounded feedback, is a promising new paradigm for automated robot design.

Jan Ludziejewski, Jan Małaśnicki, Maciej Pióro et al.

In this work, we introduce a novel approach for optimizing LLM training by adjusting learning rates across weights of different components in Transformer models. Traditional methods often apply a uniform learning rate across all network layers, potentially overlooking the unique dynamics of each part. Remarkably, our introduced relative learning rates, RLRS, method accelerates the training process by up to $23\%$, particularly in complex models such as Mixture of Experts (MoE). Hyperparameters of RLRS can be efficiently tuned on smaller models and then effectively reused on models up to $27\times$ larger. This simple and effective method results in a substantial reduction in training time and computational resources, offering a practical and scalable solution for optimizing large-scale neural networks.

Maciej Stefaniak, Michał Krutul, Jan Małaśnicki et al.

Large language models have steadily increased in size to achieve improved performance; however, this growth has also led to greater inference time and computational demands. Consequently, there is rising interest in model size reduction methods. To address this issue, we propose Projected Compression, a novel model compression technique, that reduces model weights by utilizing projection modules. Specifically, we first train additional trainable projections weights and preserve access to all the original model parameters. Subsequently, these projections are merged into a lower-dimensional product matrix, resulting in a reduced-size standard Transformer-based model. Unlike alternative approaches that require additional computational overhead, our method matches the base model's per-token computation step in FLOPs. Experimental results show that Projected Compression outperforms the comparable hard pruning and retraining approach on higher quality models. Moreover, the performance margin scales well with the number of tokens.

Michał Zawalski, Gracjan Góral, Michał Tyrolski et al.

Efficiently tackling combinatorial reasoning problems, particularly the notorious NP-hard tasks, remains a significant challenge for AI research. Recent efforts have sought to enhance planning by incorporating hierarchical high-level search strategies, known as subgoal methods. While promising, their performance against traditional low-level planners is inconsistent, raising questions about their application contexts. In this study, we conduct an in-depth exploration of subgoal-planning methods for combinatorial reasoning. We identify the attributes pivotal for leveraging the advantages of high-level search: hard-to-learn value functions, complex action spaces, presence of dead ends in the environment, or using data collected from diverse experts. We propose a consistent evaluation methodology to achieve meaningful comparisons between methods and reevaluate the state-of-the-art algorithms.

Michal Nauman, Mateusz Ostaszewski, Marek Cygan

In this paper, we investigate the issue of error accumulation in critic networks updated via pessimistic temporal difference objectives. We show that the critic approximation error can be approximated via a recursive fixed-point model similar to that of the Bellman value. We use such recursive definition to retrieve the conditions under which the pessimistic critic is unbiased. Building on these insights, we propose Validation Pessimism Learning (VPL) algorithm. VPL uses a small validation buffer to adjust the levels of pessimism throughout the agent training, with the pessimism set such that the approximation error of the critic targets is minimized. We investigate the proposed approach on a variety of locomotion and manipulation tasks and report improvements in sample efficiency and performance.

Dominik Filipiak, Piotr Tempczyk, Marek Cygan

We present n-CPS - a generalisation of the recent state-of-the-art cross pseudo supervision (CPS) approach for the task of semi-supervised semantic segmentation. In n-CPS, there are n simultaneously trained subnetworks that learn from each other through one-hot encoding perturbation and consistency regularisation. We also show that ensembling techniques applied to subnetworks outputs can significantly improve the performance. To the best of our knowledge, n-CPS paired with CutMix outperforms CPS and sets the new state-of-the-art for Pascal VOC 2012 with (1/16, 1/8, 1/4, and 1/2 supervised regimes) and Cityscapes (1/16 supervised).

Maciej Jaśkowski, Jakub Świątkowski, Michał Zając et al.

Recent developments in the field of robot grasping have shown great improvements in the grasp success rates when dealing with unknown objects. In this work we improve on one of the most promising approaches, the Grasp Quality Convolutional Neural Network (GQ-CNN) trained on the DexNet 2.0 dataset. We propose a new architecture for the GQ-CNN and describe practical improvements that increase the model validation accuracy from 92.2% to 95.8% and from 85.9% to 88.0% on respectively image-wise and object-wise training and validation splits.

Marek Cygan, Łukasz Kowalik, Arkadiusz Socała et al.

We present three results on the complexity of Minimax Approval Voting. First, we study Minimax Approval Voting parameterized by the Hamming distance $d$ from the solution to the votes. We show Minimax Approval Voting admits no algorithm running in time $\mathcal{O}^\star(2^{o(d\log d)})$, unless the Exponential Time Hypothesis (ETH) fails. This means that the $\mathcal{O}^\star(d^{2d})$ algorithm of Misra et al. [AAMAS 2015] is essentially optimal. Motivated by this, we then show a parameterized approximation scheme, running in time $\mathcal{O}^\star(\left({3}/ε\right)^{2d})$, which is essentially tight assuming ETH. Finally, we get a new polynomial-time randomized approximation scheme for Minimax Approval Voting, which runs in time $n^{\mathcal{O}(1/ε^2 \cdot \log(1/ε))} \cdot \mathrm{poly}(m)$, almost matching the running time of the fastest known PTAS for Closest String due to Ma and Sun [SIAM J. Comp. 2009].