Individual plants behave and develop differently due to variable environmental conditions (soil, local nutrient availability, local microclimate, local stress). The use of predictive (temporal) models at the individual plant scale integrated with field-level modeling using a networked, hierarchical approach will allow cyber-physical modeling and control at that plant-scale resolution for large areas. Current capabilities in persistent environmental sensing can provide hyper-local environmental conditions, suggesting that individualized models are possible. However, relying only on data to create such models leaves out contextual constraints and valuable domain knowledge. This calls for principled methods for creating individualized plant models that tightly couple multi-scale data with known biophysical and physiological knowledge. This will additionally ensure that model predictions follow known biophysical rules, thus ensuring generalizability.
Updating the individualized models require multi-modal measurements to estimate the state variables. These measurements will be performed at different scales of environmental conditions and plant physiology, and they will potentially be quite noisy due to degraded sensing environments. Hence, robust machine learning approaches are needed for feature extraction and fusion of multi-scale, multi-modal data to update the models.
Individualized actuation include localized chemical and water distribution (spraying, injection) and mechanical crop management operations. This requires dexterous actuators to enable precise chemical dispersion. Our prior work shows substantial promise in outfitting small, cheap autonomous ground vehicles with dexterous actuators in order to minimize the risk of accidents.
Projects
Our approach involves the use of advanced sensors and AI models, to accurately identify and locate weeds within a field using drones. Rather than applying herbicides across an entire area, our approach involves treatments only to the detected weeds, resulting in a substantial reduction in herbicide usage.
SRL - based Planning framework for intelligent decision makin
Conversational tool for pest identification and decision-making
This study uses multispectral UAV imagery to improve in-season yield prediction for mung bean. By analyzing key traits, it aims to identify early indicators for accurate forecasting.
A Reinforcement Learning based visual servoing policy was trained on a simulation of a soft continuum arm created in Gazebo. The policy can locate the target object in a world view camera and servo the soft arm so a tip mounted camera can view the object. Techniques such as image segmentation and controlling the arm in configuration space were used to achieve a zero-shot sim-to-real transfer.
Harvesting fruit presents many challenges due to the complex and cluttered environment. We make use of low-cost sensors such as inertial measurement units (IMUs), tactile sensors and a custom tension sensor to determine the current grasp state such as grasp stability, separation from the plant, and failed grasps. A learned Random Forest model was created and trained in the laboratory and successfully tested on real plants.
Soft Continuum Arms (SCA) show much promise, particularly in the field of agriculture, due to their compliant nature. Controlling SCA remains a significant challenge. We make use of constant curvature and torsion models to dynamically control a BR2 style SCA.
Weeds pose a significant threat to crops each year, leading to substantial economic losses for farmers. Proper identification is crucial before implementing control measures, as misidentifying a weed could result in the removal of plants that may actually benefit the crop. We propose a weed identification mobile app designed to accurately identify agriculturally important weeds. The app's classification results would also be linked with Integrated Weed Management (IWM) strategies to support effective and sustainable weed control.
Insect-pest affect crops annually and farmers incur huge economic losses as a consequence. Before coming up with measure to mitigate them, it is essential to identify them correctly, as an incorrect identification of an insect-pest could result in eradicating a pest that could prove beneficial to the crop otherwise .We propose an insect-pest identification mobile app that identifies agriculturally important insect-pests. The classification result would also be integrated with Integrated Pest Management (IPM) strategies.
Integrating Terresentia and MyCobot Elephant Arm
Attached Custom Sprayer and Camera
Incorporated MoveIt Simulation for Control and Motion Planning.
Integrating Terresentia and MyCobot Elephant Arm
Visual servoing towards a berry
Sensors collect large volumes of plant canopy reflectance data to monitor, forecast, and perform important tasks. These data are usually heterogeneous making it difficult for usage. We are exploring multimodal data fusion using cross-modal synergy to improve performance of agricultural related tasks.
This research focuses on building a simulation environment for biotic stresses with a user-definable management zone. AgGym, our proposed simulation platform, acts as a Gym environment to design deep reinforcement learning (RL) based mitigation strategies for localized plant biotic stresses without extensive field experimentation.
This research focuses on Estimating Soybean Maturity from 2D Contour Plots in UAV Time Series Imagery through ML model
Recent advances in sensor technology have enabled the cheap collection of spatial and temporal high-resolution data for agriculture across a wide geographical area. We deploy distributed deep learning algorithms for a wide variety of decision support tasks in agriculture such as as anomaly detection and image retrieval.
Developed simulation environments (“digital twins”) for robot learning tasks in agricultural settings
Created multiple realistic simulations using Blender
Each environment consists of several obstacles and targets created using geometrical meshes
Dynamic Control for quick coarse control
Visual Servoing for fine tuned manipulation
Obtaining stable images from tip camera
Soft arm controls and visual servoing
Path planning through and around obstacle
Up and coming applications like teleoperation, edge-controlled robots require high network throughput. However, existing networks are not well suited for farms either due to low throughput (TVWS, LoRa), unavailability (Cellular) or poor range (Wifi, Zigbee).
In this project, we benchmark the performance of Wifi in farms in context of high throughput applications. Our measurements show that throughput degrades quickly with distance even when using latest standard of Wifi 6. We also observe that while 5GHz band give substantially higher throughput compared to 2.4 GHz band, its range is extremely poor. Thus, there is a tradeoff between throughput and range between two bands.
Our findings will help guide us to design a traffic aware Wifi mesh network.