Alexander Kleiner

Jiajie Zhang, Sören Schwertfeger, Alexander Kleiner

We present a novel unsupervised framework to unlock vast unlabeled human demonstration data from continuous industrial video streams for Vision-Language-Action (VLA) model pre-training. Our method first trains a lightweight motion tokenizer to encode motion dynamics, then employs an unsupervised action segmenter leveraging a novel "Latent Action Energy" metric to discover and segment semantically coherent action primitives. The pipeline outputs both segmented video clips and their corresponding latent action sequences, providing structured data directly suitable for VLA pre-training. Evaluations on public benchmarks and a proprietary electric motor assembly dataset demonstrate effective segmentation of key tasks performed by humans at workstations. Further clustering and quantitative assessment via a Vision-Language Model confirm the semantic coherence of the discovered action primitives. To our knowledge, this is the first fully automated end-to-end system for extracting and organizing VLA pre-training data from unstructured industrial videos, offering a scalable solution for embodied AI integration in manufacturing.

Michal Čáp, Jiří Vokřínek, Alexander Kleiner

We consider a system consisting of multiple mobile robots in which the user can at any time issue relocation tasks ordering one of the robots to move from its current location to a given destination location. In this paper, we deal with the problem of finding a trajectory for each such relocation task that avoids collisions with other robots. The chosen robot plans its trajectory so as to avoid collision with other robots executing tasks that were issued earlier. We prove that if all possible destinations of the relocation tasks satisfy so-called valid infrastructure property, then this mechanism is guaranteed to always succeed and provide a trajectory for the robot that reaches the destination without colliding with any other robot. The time-complexity of the approach on a fixed space-time discretization is only quadratic in the number of robots. We demonstrate the applicability of the presented method on several real-world maps and compare its performance against a popular reactive approach that attempts to solve the collisions locally. Besides being dead-lock free, the presented approach generates trajectories that are significantly faster (up to 48% improvement) than the trajectories resulting from local collision avoidance.

Michal Čáp, Peter Novák, Alexander Kleiner

We deal with the problem of planning collision-free trajectories for robots operating in a shared space. Given the start and destination position for each of the robots, the task is to find trajectories for all robots that reach their destinations with minimum total cost such that the robots will not collide when following the found trajectories. Our approach starts from individually optimal trajectory for each robot, which are then penalized for being in collision with other robots. The penalty is gradually increased and the individual trajectories are iteratively replanned to account for the increased penalty until a collision-free solution is found. Using extensive experimental evaluation, we find that such a penalty method constructs trajectories with near-optimal cost on the instances where the optimum is known and otherwise with 4-10 % lower cost than the trajectories generated by prioritized planning and up to 40 % cheaper than trajectories generated by local collision avoidance techniques, such as ORCA.

Michal Čáp, Peter Novák, Alexander Kleiner et al.

An important capability of autonomous multi-robot systems is to prevent collision among the individual robots. One approach to this problem is to plan conflict-free trajectories and let each of the robots follow its pre-planned trajectory. A widely used practical method for multi-robot trajectory planning is prioritized planning, which has been shown to be effective in practice, but is in general incomplete. Formal analysis of instances that are provably solvable by prioritized planning is still missing. Moreover, prioritized planning is a centralized algorithm, which may be in many situations undesirable. In this paper we a) propose a revised version of prioritized planning and characterize the class of instances that are provably solvable by the algorithm and b) propose an asynchronous decentralized variant of prioritized planning, which maintains the desirable properties of the centralized version and in the same time exploits the distributed computational power of the individual robots, which in most situations allows to find the joint trajectories faster. The experimental evaluation performed on real-world indoor maps shows that a) the revised version of prioritized planning reliably solves a wide class of instances on which both classical prioritized planning and popular reactive technique ORCA fail and b) the asynchronous decentralized algorithm provides solution faster than the previously proposed synchronized decentralized algorithm.