arXiv subjects

Qijun He

Publications and source records attributed to Qijun He.

12 recordsLinked to original sources

arXiv2026

Can Vision-Language-Action Models Learn from Real-World Data Continually without Forgetting?

Vision-Language-Action (VLA) models provide a promising foundation for general-purpose robotics, yet their real-world deployment demands the ability to continually acquire new skills without forgetting prior ones. While recent studies have explored continual learning for VLA models in simulated settings, the challenge remains largely unexamined under realistic physical conditions. To bridge this gap, we construct a real-world continual learning benchmark comprising ten diverse sequential manipulation tasks across both single-arm and bimanual configurations. Through extensive experiments on this benchmark, we find that naive sequential fine-tuning leads to severe catastrophic forgetting, whereas a well-configured experience replay (ER) approach can effectively mitigate forgetting and outperform joint multi-task training under equivalent computational budgets. Notably, by synthesizing our empirical findings, we successfully achieve stable continual learning across the full 10-task heterogeneous stream, retaining previously acquired capabilities while adapting to diverse new skills in real-world deployment. This work presents an empirical study grounded in real-world continual VLA learning and offers actionable insights for deploying robust, long-lived robotic policies.

cs.RO↗

arXiv2026

Continual Policy Consolidation for Lifelong Robot Learning

Building a generalist robot policy requires continuously integrating new skills while preserving previously acquired behaviors. Directly optimizing a single policy over a growing task stream is difficult because robotic interaction is expensive, task distributions are heterogeneous, and sequential updates induce interference. To address these problems, we propose continual policy consolidation (CPC), a teacher--student framework that combines continual policy distillation with prioritized experience replay and expandable experts. This architecture separates skill acquisition from policy consolidation: teachers are trained independently through reinforcement learning, and their behaviors are continually distilled into a central generalist student. This decomposition retains the practical strength of reinforcement learning for task-specialized training while casting student-side consolidation as a supervised policy-learning problem. To balance stability and plasticity as the task stream grows, the student combines an expandable Transformer-based mixture-of-experts architecture with prioritized trajectory replay. Extensive experiments show that the student recovers a large proportion of teacher performance while achieving near-zero forgetting. These results demonstrate a scalable route for consolidating independently acquired robot skills into a continually growing generalist policy.

cs.LG↗

arXiv2022

The arithmetic topology of genetic alignments

We propose a novel mathematical paradigm for the study of genetic variation in sequence alignments. This framework originates from extending the notion of pairwise relations, upon which current analysis is based on, to k-ary dissimilarity. This dissimilarity naturally leads to a generalization of simplicial complexes by endowing simplices with weights, compatible with the boundary operator. We introduce the notion of k-stances and dissimilarity complex, the former encapsulating arithmetic as well as topological structure expressing these k-ary relations. We study basic mathematical properties of dissimilarity complexes and show how this approach captures an entirely new layer of biologically relevant viral dynamics in the context of SARS-CoV-2 and H1N1 flu genomic data.

q-bio.QM↗

arXiv2022

DIY-IPS: Towards an Off-the-Shelf Accurate Indoor Positioning System

We present DIY-IPS - Do It Yourself - Indoor Positioning System, an open-source real-time indoor positioning mobile application. DIY-IPS detects users' indoor position by employing dual-band RSSI fingerprinting of available WiFi access points. The app can be used, without additional infrastructural costs, to detect users' indoor positions in real time. We published our app as an open source to save other researchers time recreating it. The app enables researchers/users to (1) collect indoor positioning datasets with a ground truth label, (2) customize the app for higher accuracy or other research purposes (3) test the accuracy of modified methods by live testing with ground truth. We ran preliminary experiments to demonstrate the effectiveness of the app.

cs.NI↗

arXiv2020

Geometric characterization of data sets with unique reduced Gröbner bases

Model selection based on experimental data is an important challenge in biological data science. Particularly when collecting data is expensive or time consuming, as it is often the case with clinical trial and biomolecular experiments, the problem of selecting information-rich data becomes crucial for creating relevant models. We identify geometric properties of input data that result in a unique algebraic model and we show that if the data form a staircase, or a so-called linear shift of a staircase, the ideal of the points has a unique reduced Gro bner basis and thus corresponds to a unique model. We use linear shifts to partition data into equivalence classes with the same basis. We demonstrate the utility of the results by applying them to a Boolean model of the well-studied lac operon in E. coli.

math.AG↗

arXiv2019

Loop Homology of Bi-secondary Structures II

In this paper we further describe the features of the topological space $K(R)$ obtained from the loop nerve of $R$, for $R=(S,T)$ a bi-secondary structure. We will first identify certain distinct combinatorial structures in the arc diagram of $R$ which we will call crossing components. The main theorem of this paper shows that the total number of these crossing components equals the rank of $H_2(R)$, the second homology group of the loop nerve.

math.CO↗

arXiv2019

Loop Homology of Bi-secondary Structures

In this paper we compute the loop homology of bi-secondary structures. Bi-secondary structures were introduced by Haslinger and Stadler and are pairs of RNA secondary structures, i.e. diagrams having non-crossing arcs in the upper half-plane. A bi-secondary structure is represented by drawing its respective secondary structures in the upper and lower half-plane. An RNA secondary structure has a loop decomposition, where a loop corresponds to a boundary component, regarding the secondary structure as an orientable fatgraph. The loop-decomposition of secondary structures facilitates the computation of its free energy and any two loops intersect either trivially or in exactly two vertices. In bi-secondary structures the intersection of loops is more complex and is of importance in current algorithmic work in bio-informatics and evolutionary optimization. We shall construct a simplicial complex capturing the intersections of loops and compute its homology. We prove that only the zeroth and second homology groups are nontrivial and furthermore show that the second homology group is free. Finally, we provide evidence that the generators of the second homology group have a bio-physical interpretation: they correspond to pairs of mutually exclusive substructures.

math.GN↗

arXiv2018

Genetic robustness of let-7 miRNA sequence-structure pairs

Genetic robustness, the preservation of evolved phenotypes against genotypic mutations, is one of the central concepts in evolution. In recent years a large body of work has focused on the origins, mechanisms, and consequences of robustness in a wide range of biological systems. In particular, research on ncRNAs studied the ability of sequences to maintain folded structures against single-point mutations. In these studies, the structure is merely a reference. However, recent work revealed evidence that structure itself contributes to the genetic robustness of ncRNAs. We follow this line of thought and consider sequence-structure pairs as the unit of evolution and introduce the spectrum of inverse folding rates (IFR-spectrum) as a measurement of genetic robustness. Our analysis of the miRNA let-7 family captures key features of structure-modulated evolution and facilitates the study of robustness against multiple-point mutations.

q-bio.PE↗

arXiv2017

An efficient dual sampling algorithm with Hamming distance filtration

Recently, a framework considering RNA sequences and their RNA secondary structures as pairs, led to some information-theoretic perspectives on how the semantics encoded in RNA sequences can be inferred. In this context, the pairing arises naturally from the energy model of RNA secondary structures. Fixing the sequence in the pairing produces the RNA energy landscape, whose partition function was discovered by McCaskill. Dually, fixing the structure induces the energy landscape of sequences. The latter has been considered for designing more efficient inverse folding algorithms. We present here the Hamming distance filtered, dual partition function, together with a Boltzmann sampler using novel dynamic programming routines for the loop-based energy model. The time complexity of the algorithm is $O(h^2n)$, where $h,n$ are Hamming distance and sequence length, respectively, reducing the time complexity of samplers, reported in the literature by $O(n^2)$. We then present two applications, the first being in the context of the evolution of natural sequence-structure pairs of microRNAs and the second constructing neutral paths. The former studies the inverse fold rate (IFR) of sequence-structure pairs, filtered by Hamming distance, observing that such pairs evolve towards higher levels of robustness, i.e.,~increasing IFR. The latter is an algorithm that constructs neutral paths: given two sequences in a neutral network, we employ the sampler in order to construct short paths connecting them, consisting of sequences all contained in the neutral network.

q-bio.BM↗

arXiv2017

Small Groebner Fans of Ideals of Points

In the context of modeling biological systems, it is of interest to generate ideals of points with a unique reduced Groebner basis, and the first main goal of this paper is to identify classes of ideals in polynomial rings which share this property. Moreover, we provide methodologies for constructing such ideals. We then relax the condition of uniqueness. The second and most relevant topic discussed here is to consider and identify pairs of ideals with the same number of reduced Groebner bases, that is, with the same cardinality of their associated Groebner fan.

math.AC↗

arXiv2016

Geometric combinatorics and computational molecular biology: branching polytopes for RNA sequences

Questions in computational molecular biology generate various discrete optimization problems, such as DNA sequence alignment and RNA secondary structure prediction. However, the optimal solutions are fundamentally dependent on the parameters used in the objective functions. The goal of a parametric analysis is to elucidate such dependencies, especially as they pertain to the accuracy and robustness of the optimal solutions. Techniques from geometric combinatorics, including polytopes and their normal fans, have been used previously to give parametric analyses of simple models for DNA sequence alignment and RNA branching configurations. Here, we present a new computational framework, and proof-of-principle results, which give the first complete parametric analysis of the branching portion of the nearest neighbor thermodynamic model for secondary structure prediction for real RNA sequences.

q-bio.BM↗

arXiv2015

Stratification and enumeration of Boolean functions by canalizing depth

Boolean network models have gained popularity in computational systems biology over the last dozen years. Many of these networks use canalizing Boolean functions, which has led to increased interest in the study of these functions. The canalizing depth of a function describes how many canalizing variables can be recursively picked off, until a non-canalizing function remains. In this paper, we show how every Boolean function has a unique algebraic form involving extended monomial layers and a well-defined core polynomial. This generalizes recent work on the algebraic structure of nested canalizing functions, and it yields a stratification of all Boolean functions by their canalizing depth. As a result, we obtain closed formulas for the number of n-variable Boolean functions with depth k, which simultaneously generalizes enumeration formulas for canalizing, and nested canalizing functions.

cs.DM↗