Searcharxiv⌕ Search

arXiv subjects

Stepan Dergachev

Publications and source records attributed to Stepan Dergachev.

11 recordsLinked to original sources

Passage-Focused Soft Coordination for Hierarchical Multi-Robot Navigation in Bottleneck Environments

In multi-robot navigation, rigid global schedules are sensitive to execution deviations, while local methods lack long-horizon coordination. We propose a passage-focused framework combining soft global coordination with flexible local execution. A passage-induced graph of convex sets supports priority-ordered planning with two levels of soft coordination. Passage-aware route selection accounts for passage usage and traversal directions, while route-restricted continuous space-time planning penalizes temporal conflicts. The resulting references guide a model predictive path integral controller that avoids collisions online. Adaptive temporal and geometric tracking accommodates execution deviations, while selective replanning maintains coordination without rigid trajectory reservations. Across 1,200 instances in 4 different environments with 5-30 robots, the framework achieves a near-perfect success rate and outperforms evaluated baselines in makespan and flowtime. Hardware experiments with 8 robots demonstrate successful completion of all trials without collisions. These results highlight the effectiveness of passage-focused coordination in balancing long-horizon cooperation with flexible execution.

cs.RO↗

Does Step Law Transfer to Small-Scale Language Models? An Empirical Recalibration Below 59M Parameters

Step Law gives power-law formulas for the optimal peak learning rate eta* and batch size B* when pre-training language models. It was calibrated on models between 59M and 1B parameters; the small-model regime N < 59M was never tested empirically by its authors. This regime matters for single-GPU training, interpretability research, educational experiments, and settings where larger models are infeasible on memory or cost grounds. We test whether Step Law transfers to small language models. We consider three outcomes: H1, the original coefficients work directly; H2, the power-law form holds but with different coefficients; and H3, a power law does not describe the optima in this regime. All experiments use a single nanoGPT/TinyStories pipeline with a 2048-token BPE vocabulary, AdamW, and a warmup-cosine schedule. The optimum for each (N, D) cell is extracted from the loss surface L(eta, B) via a local quadratic approximation in log-log coordinates over the smoothed training loss. The final dataset contains 29 unique (N, D) cells and 935 analysis-ready runs. The main refit uses 25 cells (815 runs) in the working range 4 <= D/N <= 600. On the pooled data we accept H2: the functional form is preserved, but the coefficients differ from the original. We obtain eta*(N, D) = 0.0985 N^(-0.508) D^(0.238) (R^2 = 0.834) and B*(D) = 3.6 x 10^(-4) D^(0.931) (R^2 = 0.950). Step Law's structural claim that B* is independent of N is reproduced (p = 0.87), but the growth of B* with D is nearly twice as steep as in the original work. Direct transfer of Step Law systematically overestimates the optimal learning rate: the median ratio eta_SL / eta* is approximately 4.0x, with a range of 2.4x to 6.6x.

cs.LG↗

CoRL-MPPI: Enhancing MPPI With Learnable Behaviours For Efficient And Provably-Safe Multi-Robot Collision Avoidance

Decentralized collision avoidance is a core challenge for scalable multi-robot systems. A promising approach to this problem is Model Predictive Path Integral (MPPI) control - a framework that naturally handles arbitrary motion models and provides strong theoretical guarantees. Still, in practice an MPPI-based controller may produce suboptimal trajectories because its performance relies heavily on uninformed random sampling. We introduce CoRL-MPPI, a fusion of Cooperative Reinforcement Learning and MPPI that addresses this limitation. We train an action policy, approximated by a deep neural network, in simulation to learn local cooperative collision-avoidance behaviors. This learned policy is then embedded into the MPPI framework to guide its sampling distribution, biasing it toward more intelligent and cooperative actions in scenarios that may differ substantially from those used during training. Moreover, CoRL-MPPI preserves the theoretical guarantees of regular MPPI. We evaluate our approach in dense, dynamic setups against classical and learning-based state-of-the-art baselines. Our results demonstrate that CoRL-MPPI outperforms competing methods and significantly improves navigation efficiency, measured by success rate and delay, as well as safety, enabling agile and robust multi-robot navigation.

cs.RO↗

Relaxing Constraints in Anonymous Multi Agent Path Finding for Large Agents

The study addressed the problem of Anonymous Multi-Agent Path-finding (AMAPF). Unlike the classical formulation, where the assignment of agents to goals is fixed, in the anonymous MAPF setting it is irrelevant which agent reaches specific goal, provided that all goals are occupied. Most existing multi-agent pathfinding algorithms rely on a discrete representation of the environment (e.g., square grids) and do not account for the sizes of agents. This limits their applicability in real-world scenarios, such as trajectory planning for mobile robots in warehouses. Conversely, methods operating in continuous space typically impose substantial restrictions on the input data, such as constraints on the distances between initial and goal positions or between start/goal positions and obstacles. In this work, we considered one of the AMAPF algorithms designed for continuous space, where agents are modeled as disks of equal size. The algorithm requires a strict minimum separation of $4$ agent radii between any start/goal positions. Proposed a modification aimed at relaxing the constraints and reduce this limit from $4$ to $2\sqrt{3}$. We theoretically demonstrated that the proposed enhancements preserve original theoretical properties, including the guarantee that all agents will eventually achieve their goals safely and without collisions.

cs.MA↗

Decentralized Uncertainty-Aware Multi-Agent Collision Avoidance with Model Predictive Path Integral

Decentralized multi-agent navigation under uncertainty is a complex task that arises in numerous robotic applications. It requires collision avoidance strategies that account for both kinematic constraints, sensing and action execution noise. In this paper, we propose a novel approach that integrates the Model Predictive Path Integral (MPPI) with a probabilistic adaptation of Optimal Reciprocal Collision Avoidance. Our method ensures safe and efficient multi-agent navigation by incorporating probabilistic safety constraints directly into the MPPI sampling process via a Second-Order Cone Programming formulation. This approach enables agents to operate independently using local noisy observations while maintaining safety guarantees. We validate our algorithm through extensive simulations with differential-drive robots and benchmark it against state-of-the-art methods, including ORCA-DD and B-UAVC. Results demonstrate that our approach outperforms them while achieving high success rates, even in densely populated environments. Additionally, validation in the Gazebo simulator confirms its practical applicability to robotic platforms. A source code is available at http://github.com/PathPlanning/MPPI-Collision-Avoidance.

cs.RO↗

Decentralized Unlabeled Multi-Agent Navigation in Continuous Space

In this work, we study the problem where a group of mobile agents needs to reach a set of goal locations, but it does not matter which agent reaches a specific goal. Unlike most of the existing works on this topic that typically assume the existence of the centralized planner (or controller) and limit the agents' moves to a predefined graph of locations and transitions between them, in this work we focus on the decentralized scenarios, when each agent acts individually relying only on local observations/communications and is free to move in arbitrary direction at any time. Our iterative approach involves agents individually selecting goals, exchanging them, planning paths, and at each time step choose actions that balance between progressing along the paths and avoiding collisions. The proposed method is shown to be complete under specific assumptions on how agents progress towards their current goals, and our empirical evaluation demonstrates its superiority over a baseline decentralized navigation approach in success rate (i.e. is able to solve more problem instances under a given time limit) and a comparison with the centralized TSWAP algorithm reveals its efficiency in minimizing trajectory lengths for mission accomplishment.

cs.MA↗

Decentralized Unlabeled Multi-agent Pathfinding Via Target And Priority Swapping (With Supplementary)

In this paper we study a challenging variant of the multi-agent pathfinding problem (MAPF), when a set of agents must reach a set of goal locations, but it does not matter which agent reaches a specific goal - Anonymous MAPF (AMAPF). Current optimal and suboptimal AMAPF solvers rely on the existence of a centralized controller which is in charge of both target assignment and pathfinding. We extend the state of the art and present the first AMAPF solver capable of solving the problem at hand in a fully decentralized fashion, when each agent makes decisions individually and relies only on the local communication with the others. The core of our method is a priority and target swapping procedure tailored to produce consistent goal assignments (i.e. making sure that no two agents are heading towards the same goal). Coupled with an established rule-based path planning, we end up with a TP-SWAP, an efficient and flexible approach to solve decentralized AMAPF. On the theoretical side, we prove that TP-SWAP is complete (i.e. TP-SWAP guarantees that each target will be reached by some agent). Empirically, we evaluate TP-SWAP across a wide range of setups and compare it to both centralized and decentralized baselines. Indeed, TP-SWAP outperforms the fully-decentralized competitor and can even outperform the semi-decentralized one (i.e. the one relying on the initial consistent goal assignment) in terms of flowtime (a widespread cost objective in MAPF

cs.MA↗

2.5D Mapping, Pathfinding and Path Following For Navigation Of A Differential Drive Robot In Uneven Terrain

Safe navigation in uneven terrains is an important problem in robotic research. In this paper we propose a 2.5D navigation system which consists of elevation map building, path planning and local path following with obstacle avoidance. For local path following we use Model Predictive Path Integral (MPPI) control method. We propose novel cost-functions for MPPI in order to adapt it to elevation maps and motion through unevenness. We evaluate our system on multiple synthetic tests and in a simulated environment with different types of obstacles and rough surfaces.

cs.RO↗

Towards A Complete Multi-Agent Pathfinding Algorithm For Large Agents

Multi-agent pathfinding (MAPF) is a challenging problem which is hard to solve optimally even when simplifying assumptions are adopted, e.g. planar graphs (typically -- grids), discretized time, uniform duration of move and wait actions etc. On the other hand, MAPF under such restrictive assumptions (also known as the Classical MAPF) is equivalent to the so-called pebble motion problem for which non-optimal polynomial time algorithms do exist. Recently, a body of works emerged that investigated MAPF beyond the basic setting and, in particular, considered agents of arbitrary size and shape. Still, to the best of our knowledge no complete algorithms for such MAPF variant exists. In this work we attempt to narrow this gap by considering MAPF for large agents and suggesting how this problem can be reduced to pebble motion on (general) graphs. The crux of this reduction is the procedure that moves away the agents away from the edge which is needed to perform a move action of the current agent. We consider different variants of how this procedure can be implemented and present a variant of the pebble motion algorithm which incorporates this procedure. Unfortunately, the algorithm is still incomplete, but empirically we show that it is able to solve much more MAPF instances (under the strict time limit) with large agents on arbitrary non-planar graphs (roadmaps) compared to the state-of-the-art MAPF solver -- Continous Conflict-Based Search (CCBS).

cs.MA↗

Distributed Multi-agent Navigation Based on Reciprocal Collision Avoidance and Locally Confined Multi-agent Path Finding

Avoiding collisions is the core problem in multi-agent navigation. In decentralized settings, when agents have limited communication and sensory capabilities, collisions are typically avoided in a reactive fashion, relying on local observations/communications. Prominent collision avoidance techniques, e.g. ORCA, are computationally efficient and scale well to a large number of agents. However, in numerous scenarios, involving navigation through the tight passages or confined spaces, deadlocks are likely to occur due to the egoistic behaviour of the agents and as a result, the latter can not achieve their goals. To this end, we suggest an application of the locally confined multi-agent path finding (MAPF) solvers that coordinate sub-groups of the agents that appear to be in a deadlock (to detect the latter we suggest a simple, yet efficient ad-hoc routine). We present a way to build a grid-based MAPF instance, typically required by modern MAPF solvers. We evaluate two of them in our experiments, i.e. Push and Rotate and a bounded-suboptimal version of Conflict Based Search (ECBS), and show that their inclusion into the navigation pipeline significantly increases the success rate, from 15% to 99% in certain cases.

cs.MA↗

A Combination of Theta*, ORCA and Push and Rotate for Multi-agent Navigation

We study the problem of multi-agent navigation in static environments when no centralized controller is present. Each agent is controlled individually and relies on three algorithmic components to achieve its goal while avoiding collisions with the other agents and the obstacles: i) individual path planning which is done by Theta* algorithm; ii) collision avoidance while path following which is performed by ORCA* algorithm; iii) locally-confined multi-agent path planning done by Push and Rotate algorithm. The latter component is crucial to avoid deadlocks in confined areas, such as narrow passages or doors. We describe how the suggested components interact and form a coherent navigation pipeline. We carry out an extensive empirical evaluation of this pipeline in simulation. The obtained results clearly demonstrate that the number of occurring deadlocks significantly decreases enabling more agents to reach their goals compared to techniques that rely on collision-avoidance only and do not include multi-agent path planning component

cs.MA↗