Marco Bagatella

Marco Bagatella, Sammy Christen, Otmar Hilliges

Efficient exploration is a crucial challenge in deep reinforcement learning. Several methods, such as behavioral priors, are able to leverage offline data in order to efficiently accelerate reinforcement learning on complex tasks. However, if the task at hand deviates excessively from the demonstrated task, the effectiveness of such methods is limited. In our work, we propose to learn features from offline data that are shared by a more diverse range of tasks, such as correlation between actions and directedness. Therefore, we introduce state-free priors, which directly model temporal consistency in demonstrated trajectories, and are capable of driving exploration in complex tasks, even when trained on data collected on simpler tasks. Furthermore, we introduce a novel integration scheme for action priors in off-policy reinforcement learning by dynamically sampling actions from a probabilistic mixture of policy and action prior. We compare our approach against strong baselines and provide empirical evidence that it can accelerate reinforcement learning in long-horizon continuous control tasks under sparse reward settings.

Núria Armengol Urpí, Marco Bagatella, Otmar Hilliges et al.

Real-world robotic manipulation tasks remain an elusive challenge, since they involve both fine-grained environment interaction, as well as the ability to plan for long-horizon goals. Although deep reinforcement learning (RL) methods have shown encouraging results when planning end-to-end in high-dimensional environments, they remain fundamentally limited by poor sample efficiency due to inefficient exploration, and by the complexity of credit assignment over long horizons. In this work, we present Efficient Learning of High-Level Plans from Play (ELF-P), a framework for robotic learning that bridges motion planning and deep RL to achieve long-horizon complex manipulation tasks. We leverage task-agnostic play data to learn a discrete behavioral prior over object-centric primitives, modeling their feasibility given the current context. We then design a high-level goal-conditioned policy which (1) uses primitives as building blocks to scaffold complex long-horizon tasks and (2) leverages the behavioral prior to accelerate learning. We demonstrate that ELF-P has significantly better sample efficiency than relevant baselines over multiple realistic manipulation tasks and learns policies that can be easily transferred to physical hardware.

Jonas Hübotter, Frederike Lübeck, Lejs Behric et al.

Large language models are increasingly post-trained with reinforcement learning in verifiable domains such as code and math. Yet, current methods for reinforcement learning with verifiable rewards (RLVR) learn only from a scalar outcome reward per attempt, creating a severe credit-assignment bottleneck. Many verifiable environments actually provide rich textual feedback, such as runtime errors or judge evaluations, that explain why an attempt failed. We formalize this setting as reinforcement learning with rich feedback and introduce Self-Distillation Policy Optimization (SDPO), which converts tokenized feedback into a dense learning signal without any external teacher or explicit reward model. SDPO treats the current model conditioned on feedback as a self-teacher and distills its feedback-informed next-token predictions back into the policy. In this way, SDPO leverages the model's ability to retrospectively identify its own mistakes in-context. Across scientific reasoning, tool use, and competitive programming on LiveCodeBench v6, SDPO improves sample efficiency and final accuracy over strong RLVR baselines. Notably, SDPO also outperforms baselines in standard RLVR environments that only return scalar feedback by using successful rollouts as implicit feedback for failed attempts. Finally, applying SDPO to individual questions at test time accelerates discovery on difficult binary-reward tasks, achieving the same discovery probability as best-of-k sampling or multi-turn conversations with 3x fewer attempts.

Marco Bagatella, Andreas Krause, Georg Martius

Linear temporal logic (LTL) is a powerful language for task specification in reinforcement learning, as it allows describing objectives beyond the expressivity of conventional discounted return formulations. Nonetheless, recent works have shown that LTL formulas can be translated into a variable rewarding and discounting scheme, whose optimization produces a policy maximizing a lower bound on the probability of formula satisfaction. However, the synthesized reward signal remains fundamentally sparse, making exploration challenging. We aim to overcome this limitation, which can prevent current algorithms from scaling beyond low-dimensional, short-horizon problems. We show how better exploration can be achieved by further leveraging the LTL specification and casting its corresponding Limit Deterministic Büchi Automaton (LDBA) as a Markov reward process, thus enabling a form of high-level value estimation. By taking a Bayesian perspective over LDBA dynamics and proposing a suitable prior distribution, we show that the values estimated through this procedure can be treated as a shaping potential and mapped to informative intrinsic rewards. Empirically, we demonstrate applications of our method from tabular settings to high-dimensional continuous systems, which have so far represented a significant challenge for LTL-based reinforcement learning algorithms.

Marco Bagatella, Georg Martius

Curiosity has established itself as a powerful exploration strategy in deep reinforcement learning. Notably, leveraging expected future novelty as intrinsic motivation has been shown to efficiently generate exploratory trajectories, as well as a robust dynamics model. We consider the challenge of extracting goal-conditioned behavior from the products of such unsupervised exploration techniques, without any additional environment interaction. We find that conventional goal-conditioned reinforcement learning approaches for extracting a value function and policy fall short in this difficult offline setting. By analyzing the geometry of optimal goal-conditioned value functions, we relate this issue to a specific class of estimation artifacts in learned values. In order to mitigate their occurrence, we propose to combine model-based planning over learned value landscapes with a graph-based value aggregation scheme. We show how this combination can correct both local and global artifacts, obtaining significant improvements in zero-shot goal-reaching performance across diverse simulated environments.

Tim Launer, Jonas Hübotter, Marco Bagatella et al.

We investigate Functional Majority Voting (FMV), a method based on functional consensus for code generation with Large Language Models, which identifies a representative solution from multiple generations using their runtime execution signatures on test inputs. We find that FMV is an effective test-time inference strategy, substantially boosting performance on LiveCodeBench without a large compute overhead. Furthermore, we extend the utility of functional consensus and apply it as an aggregation strategy for label-free Test-Time Reinforcement Learning. We demonstrate that this increases pass@1 on holdout tasks, but find no evidence of self-improvement beyond the base model's performance ceiling.

Marco Bagatella, Thomas Rupf, Georg Martius et al.

Recent advancements in zero-shot reinforcement learning (RL) have facilitated the extraction of diverse behaviors from unlabeled, offline data sources. In particular, forward-backward algorithms (FB) can retrieve a family of policies that can approximately solve any standard RL problem (with additive rewards, linear in the occupancy measure), given sufficient capacity. While retaining zero-shot properties, we tackle the greater problem class of RL with general utilities, in which the objective is an arbitrary differentiable function of the occupancy measure. This setting is strictly more expressive, capturing tasks such as distribution matching or pure exploration, which may not be reduced to additive rewards. We show that this additional complexity can be captured by a novel, maximum entropy (soft) variant of the forward-backward algorithm, which recovers a family of stochastic policies from offline data. When coupled with zero-order search over compact policy embeddings, this algorithm can sidestep iterative optimization schemes, and optimizes general utilities directly at test-time. Across both didactic and high-dimensional experiments, we demonstrate that our method retains favorable properties of FB algorithms, while also extending their range to more general RL problems.

Thomas Rupf, Marco Bagatella, Nico Gürtler et al.

Zero-shot imitation learning algorithms hold the promise of reproducing unseen behavior from as little as a single demonstration at test time. Existing practical approaches view the expert demonstration as a sequence of goals, enabling imitation with a high-level goal selector, and a low-level goal-conditioned policy. However, this framework can suffer from myopic behavior: the agent's immediate actions towards achieving individual goals may undermine long-term objectives. We introduce a novel method that mitigates this issue by directly optimizing the occupancy matching objective that is intrinsic to imitation learning. We propose to lift a goal-conditioned value function to a distance between occupancies, which are in turn approximated via a learned world model. The resulting method can learn from offline, suboptimal data, and is capable of non-myopic, zero-shot imitation, as we demonstrate in complex, continuous benchmarks. The code is available at https://github.com/martius-lab/zilot.

Thomas Rupf, Marco Bagatella, Marin Vlastelica et al.

Behavior Foundation Models (BFMs) are capable of retrieving high-performing policy for any reward function specified directly at test-time, commonly referred to as zero-shot reinforcement learning (RL). While this is a very efficient process in terms of compute, it can be less so in terms of data: as a standard assumption, BFMs require computing rewards over a non-negligible inference dataset, assuming either access to a functional form of rewards, or significant labeling efforts. To alleviate these limitations, we tackle the problem of task inference purely through interaction with the environment at test-time. We propose OpTI-BFM, an optimistic decision criterion that directly models uncertainty over reward functions and guides BFMs in data collection for task inference. Formally, we provide a regret bound for well-trained BFMs through a direct connection to upper-confidence algorithms for linear bandits. Empirically, we evaluate OpTI-BFM on established zero-shot benchmarks, and observe that it enables successor-features-based BFMs to identify and optimize an unseen reward function in a handful of episodes with minimal compute overhead. Code is available at https://github.com/ThomasRupf/opti-bfm.

Marco Bagatella, Mert Albaba, Jonas Hübotter et al.

Foundation models compress a large amount of information in a single, large neural network, which can then be queried for individual tasks. There are strong parallels between this widespread framework and offline goal-conditioned reinforcement learning algorithms: a universal value function is trained on a large number of goals, and the policy is evaluated on a single goal in each test episode. Extensive research in foundation models has shown that performance can be substantially improved through test-time training, specializing the model to the current goal. We find similarly that test-time offline reinforcement learning on experience related to the test goal can lead to substantially better policies at minimal compute costs. We propose a novel self-supervised data selection criterion, which selects transitions from an offline dataset according to their relevance to the current state and quality with respect to the evaluation goal. We demonstrate across a wide range of high-dimensional loco-navigation and manipulation tasks that fine-tuning a policy on the selected data for a few gradient steps leads to significant performance gains over standard offline pre-training. Our goal-conditioned test-time training (GC-TTT) algorithm applies this routine in a receding-horizon fashion during evaluation, adapting the policy to the current trajectory as it is being rolled out. Finally, we study compute allocation at inference, demonstrating that, at comparable costs, GC-TTT induces performance gains that are not achievable by scaling model size.

Marco Bagatella, Matteo Pirotta, Ahmed Touati et al.

Latent prediction--where agents learn by predicting their own latents--has emerged as a powerful paradigm for training general representations in machine learning. In reinforcement learning (RL), this approach has been explored to define auxiliary losses for a variety of settings, including reward-based and unsupervised RL, behavior cloning, and world modeling. While existing methods are typically limited to single-task learning, one-step prediction, or on-policy trajectory data, we show that temporal difference (TD) learning enables learning representations predictive of long-term latent dynamics across multiple policies from offline, reward-free transitions. Building on this, we introduce TD-JEPA, which leverages TD-based latent-predictive representations into unsupervised RL. TD-JEPA trains explicit state and task encoders, a policy-conditioned multi-step predictor, and a set of parameterized policies directly in latent space. This enables zero-shot optimization of any reward function at test time. Theoretically, we show that an idealized variant of TD-JEPA avoids collapse with proper initialization, and learns encoders that capture a low-rank factorization of long-term policy dynamics, while the predictor recovers their successor features in latent space. Empirically, TD-JEPA matches or outperforms state-of-the-art baselines on locomotion, navigation, and manipulation tasks across 13 datasets in ExoRL and OGBench, especially in the challenging setting of zero-shot RL from pixels.

Kiran Doshi, Marco Bagatella, Stelian Coros

The combination of behavioural cloning and neural networks has driven significant progress in robotic manipulation. As these algorithms may require a large number of demonstrations for each task of interest, they remain fundamentally inefficient in complex scenarios, in which finite datasets can hardly cover the state space. One of the remaining challenges is thus out-of-distribution (OOD) generalisation, i.e. the ability to predict correct actions for states with a low likelihood with respect to the state occupancy induced by the dataset. This issue is aggravated when the system to control is treated as a black-box, ignoring its physical properties. This work characterises widespread properties of robotic manipulation, specifically pose equivariance and locality. We investigate the effect of the choice of problem space on OOD performance of BC policies and how transformations arising from characteristic properties of manipulation could be employed for its improvement. We empirically demonstrate that these transformations allow behaviour cloning policies, using either standard MLP-based one-step action prediction or diffusion-based action-sequence prediction, to generalise better to OOD problem instances.

Marco Bagatella, Mirek Olšák, Michal Rolínek et al.

The ability to form complex plans based on raw visual input is a litmus test for current capabilities of artificial intelligence, as it requires a seamless combination of visual processing and abstract algorithmic execution, two traditionally separate areas of computer science. A recent surge of interest in this field brought advances that yield good performance in tasks ranging from arcade games to continuous control; these methods however do not come without significant issues, such as limited generalization capabilities and difficulties when dealing with combinatorially hard planning instances. Our contribution is two-fold: (i) we present a method that learns to represent its environment as a latent graph and leverages state reidentification to reduce the complexity of finding a good policy from exponential to linear (ii) we introduce a set of lightweight environments with an underlying discrete combinatorial structure in which planning is challenging even for humans. Moreover, we show that our methods achieves strong empirical generalization to variations in the environment, even across highly disadvantaged regimes, such as "one-shot" planning, or in an offline RL paradigm which only provides low-quality trajectories.