ORCHARDBENCH: 物理约束的 GPU 并行苹果园仿真基准

Jul 16, 2026

Upload by Riley

Robotic tree-fruit harvesting remains a flagship challenge for agricultural automation, slowed by costly, irreproducible field experiments where each orchard tree is unique and failures can damage crops and plants. ORCHARDBENCH delivers a physically grounded, GPU-parallel apple-orchard simulation on NEWTON/MuJoCo-Warp, using stochastic L-systems to grow fully articulated trees with beam-theory-based compliant torsional joints, load- and rupture-aware branch breaking, and fruit detachment under literature-grounded pull forces. A controllable foliage occlusion layer and a depth-camera harvesting loop run at interactive rates on a single 8 GB laptop GPU, with metrics spanning completeness, throughput, and plant damage.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

Abstract

Introduction

ORCHARDBENCH: A Physically-Grounded, GPU-Parallel Apple-Orchard Simulation Benchmark for Agricultural Robotics Humphrey Munn School of EECS, The University of Queensland [h.munn@uq.net.au](h.munn@uq.net.au) Project page & code: [https://humphreymunn.github.io/orchardbench-page/](https://humphreymunn.github.io/orchardbench-page/) arXiv :2607 .06337v 1 [ cs .RO] 7 Jul 2026 Fig. 1: ORCHARDBENCH. A physically-grounded, GPU-parallel apple orchard. Left: a mobile manipulator (Clearpath Ridgeback base and Franka arm) autonomously harvesting a compliant, fruit-bearing tree over bumpy terrain—here holding a picked apple beside the collection bucket, amid foliage that moves with the branches. Right: one hundred domain-randomized trees simulatedin parallel, each a distinct plausible tree grown by a stochastic L-system, varying in growth habit, foliage, fruit and colour. Branches snap under load and apples detach at realistic pull forces (Section IV) . Abstract—Robotic tree-fruit harvesting is a flagship problem for agricultural automation, but progress is bottlenecked by the cost and irreproducibility of field experiments: an orchard is available only weeks a year, every tree is different, and a control error can permanently damage the crop or the plant. The tree models used in graphics and agronomy are geometrically detailed but physically inert, while the GPU-parallel simulators used in robot learning contain no plausible trees. We present ORCHARDBENCH, a physically-grounded, GPU-parallel simulation of apple-orchard trees on the NEWTON/MuJoCo-Warp engine. Each tree is grown by a stochastic L-system and instantiated as a fully articulated body: branches are compliant torsional spring-dampers whose stiffness follows Euler-Bernoulli beam theory, they break at a wood modulus of rupture and fall as free hinges, and apples are independent bodies on stem tethers that detach at literaturegrounded pull forces and load the branch as they are pulled. A moving, density-controllable foliage layer occludes the canopy as real leaves do. Every physical parameter is tied to a published source. Per-environment domain randomization makes each batched world a distinct tree, and a mobile manipulator with a wrist depth camera closes the loop with geometric fruit perception and an autonomous harvesting baseline. Careful engineering of the solver and the model lets ORCHARDBENCH run many parallel environments at interactive rates on a single 8 GB laptop GPU. We define the tasks and a metric suite spanning harvest completeness, throughput, and plant damage (with a per-canopyzone breakdown), and report baseline results across foliage, fruit load, terrain, canopy zone, and parallelism. The analytic baseline succeeds on about 40 % of the fruit it detects and harvests only about an eighth of the reachable fruit on a tree, leaving clear headroom for the learned and agentic methods the benchmark is built to measure. Code and videos are released at [https://humphreymunn.github.io/orchardbench-page/](https://humphreymunn.github.io/orchardbench-page/). Index Terms—agricultural robotics, physics simulation, benchmark, fruit harvesting, domain randomization, manipulation, sim-to-real. I. INTRODUCTION Tree-fruit harvesting is one of the most labour-intensive and least automated operations in agriculture. Apples are still picked overwhelmingly by hand, the seasonal workforce is shrinking and expensive, and the picking window for a given block is only a few weeks long. These pressures have driven two decades of work on robotic harvesters [1–3], but the problem remains open: fruit is heavily occluded by foliage and by the plant’sown structure, the manipulator must reach into a compliant canopy without snapping branches or bruising fruit, and every tree is geometrically unique. Crucially, the very thing that makes the task hard (physical contact with a living, breakable plant) also makes it expensive to study. A field trial requires areal orchard in se

FAQs

为什么需要 ORCHARDBENCH 这样的仿真基准？

田间实验受限于季节与可用性，且树木差异和控制误差会导致不可重复与潜在的作物/植株损害。仿真能够在更高规模上研究接触与可破坏植物的操控过程。

ORCHARDBENCH 的物理建模有哪些关键特征？

树由随机 L-system 生长，作为全关节动力学系统建模；枝干用符合欧拉-伯努利梁理论的扭转弹簧阻尼表示顺应性，超过材料断裂模量的弯曲力矩会触发断裂，并以真实接触与受力条件处理枝干与果实脱离。

该基准如何实现规模化并行与评测？

通过每个环境的域随机化生成批量且各不相同的树，使 GPU 可并行运行大量场景；同时给出覆盖采摘完整度、吞吐量与植物损伤的指标，并在多种场景因素（如遮挡、果实负载、地形与树冠分区）上报告基线结果。

Jul 16, 2026

Upload by Riley