Daniil Zelezetsky

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/.

Egor Cherepanov, Nikita Kachaev, Daniil Zelezetsky et al.

Vision-language-action (VLA) models predict chunks of future actions from the current observation, an assumption that fails under partial observability, where decisions depend on information no longer visible. Existing memory-augmented VLAs simultaneously introduce recurrence, retrieval, compression modules, auxiliary objectives, hierarchical memory, or task-specific architectural changes, so the contribution of recurrence itself remains entangled with surrounding machinery. We present a controlled isolation study of recurrence in a strong pretrained VLA backbone. Our formulation augments the transformer with a small set of learnable memory tokens carried across timesteps and updated through self-attention, trained end to end with truncated backpropagation through time, with no auxiliary losses and no architectural changes. We instantiate this as $μ$VLA, a family of OpenVLA-OFT variants parameterized by memory width m, TBPTT length K, and the memory update rule (cross-step gradients or a detached EMA), so that recurrence is the only varying factor. On MIKASA-Robo, $μ$VLA improves average success rate on five training tasks from 0.42 to 0.84 at the strongest setting and reaches 0.23 on held-out tasks with the same memory structure versus 0.07 for the memoryless baseline. On tasks requiring different memory structure, performance remains near baseline. On LIBERO, the strongest recurrent variant achieves 96.2% average success, indicating no regression under full observability. We interpret these results as a calibration of the capability envelope of minimal in-backbone recurrence, identifying the regime in which it is sufficient and the regime where additional memory structure is required. Demos and videos can be found in https://avanturist322.github.io/mu-vla/.

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.

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.