Search

This work proposes a robotic pipeline for picking and constrained placement of objects without geometric shape priors. Compared to recent efforts developed for similar tasks, where every object was assumed to be novel, the proposed system recognizes previously manipulated objects and performs online model reconstruction and reuse. Over a lifelong manipulation process, the system keeps learning features of objects it has interacted with and updates their reconstructed models. Whenever an instance of a previously manipulated object reappears, the system aims to first recognize it and then register its previously reconstructed model given the current observation. This step greatly reduces object shape uncertainty allowing the system to even reason for parts of objects, which are currently not observable. This also results in better manipulation efficiency as it reduces the need for active perception of the target object during manipulation. To get a reusable reconstructed model, the proposed pipeline adopts: i) TSDF for object representation, and ii) a variant of the standard particle filter algorithm for pose estimation and tracking of the partial object model. Furthermore, an effective way to construct and maintain a dataset of manipulated objects is presented. A sequence of real-world manipulation experiments is performed. They show how future manipulation tasks become more effective and efficient by reusing reconstructed models of previously manipulated objects, which were generated during their prior manipulation, instead of treating objects as novel every time.

Tensegrity robots, which are composed of compressive elements (rods) and flexible tensile elements (e.g., cables), have a variety of advantages, including flexibility, low weight, and resistance to mechanical impact. Nevertheless, the hybrid soft-rigid nature of these robots also complicates the ability to localize and track their state. This work aims to address what has been recognized as a grand challenge in this domain, i.e., the state estimation of tensegrity robots through a markerless, vision-based method, as well as novel, onboard sensors that can measure the length of the robot's cables. In particular, an iterative optimization process is proposed to track the 6-DoF pose of each rigid element of a tensegrity robot from an RGB-D video as well as endcap distance measurements from the cable sensors. To ensure that the pose estimates of rigid elements are physically feasible, i.e., they are not resulting in collisions between rods or with the environment, physical constraints are introduced during the optimization. Real-world experiments are performed with a 3-bar tensegrity robot, which performs locomotion gaits. Given ground truth data from a motion capture system, the proposed method achieves less than 1~cm translation error and 3 degrees rotation error, which significantly outperforms alternatives. At the same time, the approach can provide accurate pose estimation throughout the robot's motion, while motion capture often fails due to occlusions.

Tensegrity robots, composed of rigid rods and flexible cables, exhibit high strength-to-weight ratios and significant deformations, which enable them to navigate unstructured terrains and survive harsh impacts. They are hard to control, however, due to high dimensionality, complex dynamics, and a coupled architecture. Physics-based simulation is a promising avenue for developing locomotion policies that can be transferred to real robots. Nevertheless, modeling tensegrity robots is a complex task due to a substantial sim2real gap. To address this issue, this paper describes a Real2Sim2Real (R2S2R) strategy for tensegrity robots. This strategy is based on a differentiable physics engine that can be trained given limited data from a real robot. These data include offline measurements of physical properties, such as mass and geometry for various robot components, and the observation of a trajectory using a random control policy. With the data from the real robot, the engine can be iteratively refined and used to discover locomotion policies that are directly transferable to the real robot. Beyond the R2S2R pipeline, key contributions of this work include computing non-zero gradients at contact points, a loss function for matching tensegrity locomotion gaits, and a trajectory segmentation technique that avoids conflicts in gradient evaluation during training. Multiple iterations of the R2S2R process are demonstrated and evaluated on a real 3-bar tensegrity robot., Comment: Submitted to IROS2023

Soil microbial life in vivo is a major component of soils throughout the whole process of soil development.The living environment for soil microorganisms is a very sensitive early-warning indicator of change in soil ecosystems.Using slope cropland as a contrast,we investigated the main soil microbial populations,microbial biomass carbon(MBC),nitrogen(MBN),and phosphorus(MBP) and their fractal characteristic model,as well as soil enzyme activities under seven grain-for-green patterns(cultivation of Zenia insignis,Toona sinensis,Castanea mollissima,Citrus reticulate,Zenia insignis + Guimu No.1 forage,Guimu No.1 forage,and abandoned cropland) in depressions between karst hills.The main soil microbial populations differed under different grain-for-green patterns.Actinomycetes were 71.28% to 96.44% of the microbial population in the slope cropland,Zenia insignis + Guimu No.1,Guimu No.1,and abandoned cropland.Differences in their populations covered up to three orders of magnitude.Bacteria dominated in the other four soils,being 55.24% to 71.79% of the microbial population,but fungi were less than 1%.Bacteria were highest under Guimu No.1 and abandoned cropland,and were lowest in the slope cropland.Fungi populations were in the order: abandoned cropland Zenia insignis Guimu No.1 Citrus reticulate Castanea mollissima slope cropland Toona sinensis Zenia insignis + Guimu No.1,with significant differences between them.All the grain-for-green patterns except Zenia insignis increased total soil microbial populations.The differences between Guimu No.1,abandoned cropland,and Zenia insignis + Guimu No.1 were significant(P0.05),and were significantly higher than the others.Microbial biomass carbon(MBC) in the all grain-for-green patterns except Zenia insignis was significantly higher than in the slope cropland,and was highest in the abandoned cropland.Microbial biomass nitrogen(MBN) varied within a small range,showing a downward trend.Differences in MBN between cultivation of Zenia insignis + Guimu,Citrus reticulata,Zenia insignis and slope cropland were significant(P0.05).Microbial biomass phosphorus(MBP) varied widely in Toona sinensis,Castanea mollissima,abandoned cropland and Guimu No.1 but was significantly higher than in slope cropland.MBC was well correlated with bacteria(D=-4.07,R=0.81,P0.01),followed by the correlation between MBC and actinomycetes(D=3.82,R=0.44,P0.05) and that between MBN and fungi(D=0.58,R=0.61,P0.01).However,there was no fractal relationship between the other microbial biomass and microbial populations.Compared with the slope cropland,all the grain-for-green patterns except Guimu No.1 forage significantly increased the activities of soil sucrase,urease and phosphatase,suggesting an increase in carbon and nitrogen cycling.Under different grain-for-green patterns,the activity of soil sucrase was highest under Toona sinensis,reaching(29.10±7.73) mg g-1 d-1,while the activity of soil urease was highest in the abandoned cropland,reaching(540.7±34.77) μg g-1 d-1,and the activity of soil phosphatase was highest under Castanea mollissima,reaching(991.1±76.28) μg g-1 d-1.All the grain-for-green patterns improved soil microbial properties,suggesting important benefits arising from fragile ecosystem restoration and reconstruction in depressions between karst hills.

Soil microbe is the most important decomposer in forest ecosystems,and plays a key role in material recycling and energy conversion.In this study,the spatial patterns of soil microbial biomass in typical Karst peak-cluster depressions were analyzed through both geo-statistical and traditional statistical methods.The results showed that large variation existed in soil microbial biomass.Soil microbial biomass carbon(Cmic),soil microbial biomass nitrogen(Nmic),and soil microbial biomass phosphorus(Pmic) varied from 44.29—5209.63,20.91—1894.37,0.34—77.06 mg/kg,respectively.The Cmic was significantly positively correlated with Nmic.The value of Cmic/Nmic is 4.78,which obviously lower than that in other ecosystems.Semivariance analysis revealed that the Gaussian model fitted best for Cmic and Nmic,the Spherical model fitted best for Pmic,and the Exponential model fitted best for Cmic/Nmic.The SH%(C/(C+C0)×100%) of soil microbial biomass exhibited moderate spatial autocorrelation,ranging from 25% to 75%.Therefore,the spatial patterns were affected by both random and structure factors.The autocorrelation distance of Cmic and Nmic were about 50 m.Moran′s I decreased gradually with the increase of separation distance,and showed a statistically negative correlation.Moran′s I of Pmic increased when the separation distance was larger than 70 m,which indicated positive correlation.Moran′s I of Cmic/Nmic was lower than others,ranging from-0.2 to 0.2 m.The spatial distribution of Cmic and Nmic were similar,and showed a convex-type distribution in which the contents were high in the middle and low on the two sides.In contrast,The Pmic was distributed with a pattern of high in the upper slope and depressions.The content of Cmic/Nmic showed a concave-type distribution with many small patches.We conclude that soil microbes exist in a certain spatial pattern,and that its contents sharply reduced after disturbing.As a result,the protection of Karst primary ecosystems against human disturbing should be strengthened.

The spatial variability of surface soil(0—15 cm) organic matter in plantation,secondary forest,and primary forest in depressions between hills in a karst region was examined using farmland as a control.The ecological processes and mechanisms behind this variability were also discussed.Eighty sample plots of 10 m × 10 m were established in 200 m × 40 m farmland,plantation,secondary fores,and primary forest plots in depressions between karst hills.Geostatistics was used to analyze the spatial pattern of surface soil organic matter in the plots and principal component analysis and correlation analysis were used to analyze the relationships with other factors.The soil organic matter content in the depressions between karst hills was high.Along the restoration gradient from farmland plantation secondary forest primary forest,the surface soil organic matter content significantly increased and the coefficients of variation also increased.The vegetation in primary forest was well preserved and soil organic matter was up to 118 g/kg,3.76 times that of farmland.The coefficients of variation of soil organic matter in the farmland and three forest types were 19.4%—48.5%.There was a fine spatial autocorrelation in the surface soil organic matter in the farmland and the three forest types.The farmland and plantation forest were strongly influenced by humans and therefore more balanced.This meant the correlogram range was large and the maximum correlogram coefficient Moran′s I was 0.460 in the farmland and 0.780 in the plantation.The natural restoration time of the secondary forest was 22 years.Here there were more,but unevenly distributed,vegetation types meaning the correlogram range was smaller and Moran′s I coefficients fluctuated considerably.In the primary forest,however,the disturbance was low and vegetation intact.This meant the correlation was mainly affected by the topography.The best fitting models for semi-variation of secondary forest soil organic matter function in Karst peak-cluster depressions are the exponential model and the Gaussian model.The resulting R2 values of 0.926—0.971 demonstrate how well they reflected the soil organic matter spatial structure characteristics.The value of C0/(C0+C) of the surface soil organic matter in farmland was 26.5%,indicating a medium spatial correlation.The values of C0/(C0+C) in the three forest types ranged from 9.0% to 22.6%,suggesting strong spatial correlations.The spatial pattern of surface soil organic matter in farmland and plantation presented a unimodal distribution: in secondary forest it had a concave distribution and in primary forest it had a convex distribution.Soil nutrient content was the largest influencing factor on the variation in the farmland,topography and species diversity were the largest influencing factors in the plantation and primary forest,and forest structure and species diversity were the largest influencing factors in the secondary forest.Even when considering the same factor in the three forest types,the functions and correlations differed.Therefore,the corresponding strategies of fixing carbon should take the spatial variability of surface soil organic matter and its largest influencing factors in farmland and the three forest types into account.

Large scale visual search has been one of the challenging issues in the era of big data. It demands techniques that are not only highly effective and efficient but also allow users conveniently express their information needs and refine their intents. In this thesis, we focus on developing an exploratory framework for large scale visual search. We also develop a number of enabling techniques in this thesis, including compact visual content representation for scalable search, near duplicate video shot detection, and action based event detection. We propose a novel scheme for extremely low bit rate visual search, which sends compressed visual words consisting of vocabulary tree histogram and descriptor orientations rather than descriptors. Compact representation of video data is achieved through identifying keyframes of a video which can also help users comprehend visual content efficiently. We propose a novel Bag-of-Importance model for static video summarization. Near duplicate detection is one of the key issues for large scale visual search, since there exist a large number nearly identical images and videos. We propose an improved near-duplicate video shot detection approach for more effective shot representation. Event detection has been one of the solutions for bridging the semantic gap in visual search. We particular focus on human action centred event detection. We propose an enhanced sparse coding scheme to model human actions. Our proposed approach is able to significantly reduce computational cost while achieving recognition accuracy highly comparable to the state-of-the-art methods. At last, we propose an integrated solution for addressing the prime challenges raised from large-scale interactive visual search. The proposed system is also one of the first attempts for exploratory visual search. It provides users more robust results to satisfy their exploring experiences.