Searcharxiv⌕ Search

arXiv subjects

Po-Yi Wu

Publications and source records attributed to Po-Yi Wu.

9 recordsLinked to original sources

Correct First, Fast Second: What a Local 7B Code Model Can and Cannot Do for GPU Kernel Optimization

We study how much the interaction strategy around a fixed, locally hosted 7B-parameter code model (Qwen2.5-Coder-7B-Instruct) affects its ability to replace PyTorch operators with custom CUDA kernels. On 30 KernelBench tasks (levels 1-3, first 10 each, evaluated on a V100), we compare four regimes of increasing structure: zero-shot prompting, profiling-guided prompting, iterative refinement on execution feedback, and a multi-agent loop (analysis, retrieval, generation, evaluation, feedback triage) in which the same model fills four roles around a non-LLM evaluator. Three findings stand out. (1) Correctness, not speed, is the binding constraint: even correct generated kernels average 0.97-0.99x the speed of eager PyTorch, because a 7B model cannot out-write cuBLAS/cuDNN on dense GEMM and convolution; the only real wins are fused memory-bound epilogues. (2) The best strategy depends on difficulty: profiling-guided prompting is strongest overall (43% correct, 27% faster than baseline), but the agent loop is the only method that solves any level-2 fusion problem (3/10 vs. 0/10 for all others). (3) Naive agentic scaffolding is actively harmful: our first agent scored 3% correct, below zero-shot's 30%, and the decisive repairs were matching the agent's ambition to the model's competence and spending compute on best-of-n sampling under a real execution oracle (3% to 33% correct from n=1 to n=8), not on more reasoning steps.

cs.LG↗

CRONOS: Benchmarking Non-Episodic Reinforcement Learning for Robotic Manipulation

Reinforcement learning (RL) is promising for adapting robot policies to unstructured real-world environments. However, standard RL pipelines rely on episodic training with frequent scene resets, which is impractical for real-world deployment due to the need for substantial human intervention. While prior work has studied non-episodic RL in relatively small policies trained from scratch, its behavior when fine-tuning modern large-scale pre-trained vision-language-action (VLA) models remains underexplored. In this work, we introduce CRONOS, a simulation benchmark for RL fine-tuning of VLAs under non-episodic scene resets. CRONOS leverages high-fidelity physics simulators and adopts shared-scene multi-task settings. Using CRONOS, we show that naively fine-tuning pre-trained policies fails in non-episodic settings due to shifts in object configurations and end-effector poses; however, these challenges can be mitigated through targeted reset allocation and by addressing biases in pre-trained models. Moreover, we find that non-episodic training enhances long-horizon manipulation and improves generalization to previously unseen object configurations and task sequences. Project page: https://embodiedai-ntu.github.io/cronos/

cs.RO↗

Numerical Stability of Extended-Domain Spectral Collocation: Floating-Point Regularization and an Equilibrium Law for the Frame Constants

Extended-domain spectral collocation expands the solution of a boundary value problem in the Laplacian eigenbasis of an enclosing box and collocates at as many nodes as modes. The square system is exponentially ill-conditioned, and the impossibility theorem for equispaced data rules out stable geometric convergence; yet such solvers converge geometrically to a few orders of magnitude above the unit roundoff. We explain why in one dimension. The collocation solution is the solution operator applied to the interpolant of the data in an operator-dependent test space whose Lebesgue constant must grow exponentially for geometric convergence, so exact solutions can amplify data errors exponentially. In floating-point arithmetic the error of any coefficient vector is bounded by its nodal residual, its distance to a competitor and two frame constants, a leakage and a restricted Marcinkiewicz--Zygmund constant, bounded uniformly in the number of modes at a fixed positive threshold; for a backward-stable solve, under a condition on the perturbed matrix, the bound is uniform in the conditioning for suitably analytic data and, with computed constants, within a factor of about ten of the error floor of a direct solver. For the Poisson operator the rate of the two constants along the tail of the spectrum is given by a constrained equilibrium problem of logarithmic potential theory. For an interval of length $L$ extended by $δ$ at each end, the problems for $δ$ and $L^2/(4δ)$ are dual, the rate is linear for $δ= L/2$, and it starts with slope minus the fraction of the extended domain occupied by the extension, so for $δ< L/2$ the leakage constant obeys no power law with the exponent of the end of the spectrum. In experiments, convection--diffusion operators behave like the Poisson operator at moderate Péclet numbers.

math.NA↗

On the Pre-Asymptotic Stability and Inverse Structure of Extended-Domain Spectral Methods

The extended-domain method is a strategy for applying spectral methods to complex geometries. Its stability is complicated by the ill-conditioning of the Fourier extension frame. This paper provides an analysis of the method's pre-asymptotic behavior. We confirm that the spectral collocation system is asymptotically ill-conditioned for both the Poisson and convection-diffusion operators, driven by the redundancy of the underlying frame. However, we prove a fundamental structural dichotomy in their discrete Green's functions. We show that the inverse of the convection-diffusion operator is numerically highly asymmetric, exhibiting exponential upstream decay, in contrast to the numerically dense inverse of the Poisson operator. This intrinsic asymmetry explains why the convection-diffusion operator is significantly more robust to the underlying frame instability in practical computations.

math.NA↗

TaRL: Learning General and Physical Rewards from Tactile Demonstrations

Contact-rich manipulation requires robots to sequence precise contacts, maintain stable grasps, and apply directed forces. Reinforcement learning (RL) can acquire such behaviors automatically, but its performance hinges on reward design: sparse rewards reduce the learning efficiency, while dense rewards are hard to specify. Visual reward learning addresses this by inferring rewards from action-free demonstrations. Because it conditions only on visual observations, it fails to capture rewards beyond visual goals. We propose Tactile Reward Learning (TaRL), a framework that learns rewards from tactile demonstrations. TaRL takes a sequence of tactile deformation maps as input, and regresses task-completion progress from both successful and failed demonstrations. Because TaRL captures local robot-object interaction, it provides informative feedback to learn firm grasps and correctly directed forces; meanwhile, it is robust to changes in scene layout such as object position. We evaluate TaRL on four manipulation tasks in simulation and two in the real world. Used as a shaping reward, it substantially improves both sample efficiency and final success rate, raising success on Nut threading from 34% to 56% in simulation and on cube pickup from 37% to 97% in the real world. Combining tactile with visual rewards improves performance further. TaRL also generalizes across object instances: trained on box placement and directly deployed to can placement, it significantly improves policy learning on the new task. Project page is available at https://embodiedai-ntu.github.io/tarl.

cs.RO↗

MVA 2025 Small Multi-Object Tracking for Spotting Birds Challenge: Dataset, Methods, and Results

Small Multi-Object Tracking (SMOT) is particularly challenging when targets occupy only a few dozen pixels, rendering detection and appearance-based association unreliable. Building on the success of the MVA2023 SOD4SB challenge, this paper introduces the SMOT4SB challenge, which leverages temporal information to address limitations of single-frame detection. Our three main contributions are: (1) the SMOT4SB dataset, consisting of 211 UAV video sequences with 108,192 annotated frames under diverse real-world conditions, designed to capture motion entanglement where both camera and targets move freely in 3D; (2) SO-HOTA, a novel metric combining Dot Distance with HOTA to mitigate the sensitivity of IoU-based metrics to small displacements; and (3) a competitive MVA2025 challenge with 78 participants and 308 submissions, where the winning method achieved a 5.1x improvement over the baseline. This work lays a foundation for advancing SMOT in UAV scenarios with applications in bird strike avoidance, agriculture, fisheries, and ecological monitoring.

cs.CV↗

Convergent Operator-Splitting Scheme for Viscosity Solutions: A Foundation for Learning Domain-to-Solution Maps

This work introduces and rigorously analyzes a novel operator-splitting finite element scheme for approximating viscosity solutions of a broad class of constrained second-order partial differential equations. By decoupling the primary PDE evolution from the enforcement of constraints, the proposed method combines a stabilized finite element method for spatial discretization with an efficient semi-implicit time-stepping strategy. The cornerstone of our analysis is a proof that the scheme satisfies a discrete comparison principle. We demonstrate that under a mild time-step restriction and with appropriate stabilization, the discrete operator yields an M-matrix, which is sufficient to guarantee the scheme's monotonicity and consequent $L^\infty$-stability. These properties -- consistency, stability, and monotonicity -- are shown to be sufficient to prove convergence of the numerical approximation to the unique viscosity solution within the celebrated Barles--Souganidis framework. For solutions with enhanced regularity, we further establish an optimal-order error estimate of $O(Δt + h^2)$. The rigorously established stability of the scheme provides a blueprint for a novel Physics-Constrained Neural Operator (PCNO) architecture. We prove that by emulating the scheme's structure, the PCNO can provably break the curse of dimensionality for the challenging class of domain-to-solution mapping problems with complex topological variations, a problem for which standard learning approaches often fail. Numerical experiments for both a Hamilton-Jacobi equation with state constraints and a controlled reaction-diffusion system are presented to validate the theoretical findings and demonstrate the scheme's effectiveness.

math.NA↗

A Projection Method for Compressible Generic Two-Fluid Model

A new projection method for a generic two-fluid model is presented in this work. Specifically, we extend the projection method, originally designed for single-phase variable density incompressible and compressible flows, to viscous compressible two-fluid flows. The key idea is that the single pressure $p$ can be uniquely determined by the products of volume fractions and densities $ϕ_k ρ_k$, of the two fluids. Additionally, the stability of the method is ensured by appropriately assigning intermediate step variables at the time-discrete level and incorporating a stabilizing term at the fully discrete level. We prove the energy stability of the proposed numerical scheme, and its validity is demonstrated through three numerical tests.

math.NA↗

Development of a New Spectral Collocation Method Using Laplacian Eigenbasis for Elliptic Partial Differential Equations in an Extended Domain

The recent development of spectral method has been praised for its high-order convergence in simulating complex physical problems. The combination of embedded boundary method and spectral method becomes a mainstream way to tackle geometrically complicated problems. However, the convergence is deteriorated when embedded boundary strategies are employed. Owing to the loss of regularity, in this paper we propose a new spectral collocation method which retains the regularity of solutions to solve differential equations in the case of complex geometries. The idea is rooted in the basis functions defined in an extended domain, which leads to a useful upper bound of the Lebesgue constant with respect to the Fourier best approximation. In particular, how the stretching of the domain defining basis functions affects the convergence rate directly is detailed. Error estimates chosen in our proposed method show that the exponential decay convergence for problems with analytical solutions can be retained. Moreover, two-dimensional Poisson equations and convection-diffusion equations with simple and complex geometrical domains will be simulated. The predicted results justify the advantages of applying our method to tackle geometrically complicated problems.

math.NA↗