The widespread use of herbicides in cropping systems has led to the evolution of resistance in major weeds. The resultant loss of herbicide efficacy is compounded by a lack of new herbicide sites of action, driving demand for alternative weed control technologies. While there are many alternative methods for control, identifying the most appropriate method to pursue for commercial development has been hampered by the inability to compare techniques in a fair and equitable manner. Given that all currently available and alternative weed control methods share an intrinsic energy consumption, the aim of this review was to compare methods based on energy consumption. Energy consumption was compared for chemical, mechanical, and thermal weed control technologies when applied as broadcast (whole-field) and site-specific treatments. Tillage systems, such as flex-tine harrow (4.2 to 5.5 MJ ha−1), sweep cultivator (13 to 14 MJ ha−1), and rotary hoe (12 to 17 MJ ha−1) consumed the least energy of broadcast weed control treatments. Thermal-based approaches, including flaming (1,008 to 4,334 MJ ha−1) and infrared (2,000 to 3,887 MJ ha−1), are more appropriate for use in conservation cropping systems; however, their energy requirements are 100- to 1,000-fold greater than those of tillage treatments. The site-specific application of weed control treatments to control 2-leaf-stage broadleaf weeds at a density of 5 plants m−2 reduced energy consumption of herbicidal, thermal, and mechanical treatments by 97%, 99%, and 97%, respectively. Significantly, this site-specific approach resulted in similar energy requirements for current and alternative technologies (e.g., electrocution [15 to 19 MJ ha−1], laser pyrolysis [15 to 249 MJ ha−1], hoeing [17 MJ ha−1], and herbicides [15 MJ ha−1]). Using similar energy sources, a standardized energy comparison provides an opportunity for estimation of weed control costs, suggesting site-specific weed management is critical in the economically realistic implementation of alternative technologies.

No abstract

Offline reinforcement learning (RL) aims to find near-optimal policies from logged data without further environment interaction. Model-based algorithms, which learn a model of the environment from the dataset and perform conservative policy optimisation within that model, have emerged as a promising approach to this problem. In this work, we present Robust Adversarial Model-Based Offline RL (RAMBO), a novel approach to model-based offline RL. To achieve conservatism, we formulate the problem as a two-player zero sum game against an adversarial environment model. The model is trained minimise the value function while still accurately predicting the transitions in the dataset, forcing the policy to act conservatively in areas not covered by the dataset. To approximately solve the two-player game, we alternate between optimising the policy and optimising the model adversarially. The problem formulation that we address is theoretically grounded, resulting in a PAC performance guarantee and a pessimistic value function which lower bounds the value function in the true environment. We evaluate our approach on widely studied offline RL benchmarks, and demonstrate that our approach achieves state of the art performance.

We consider robot learning in the context of shared autonomy, where control of the system can switch between a human teleoperator and autonomous control. In this setting we address reinforcement learning, and learning from demonstration, where there is a cost associated with human time. This cost represents the human time required to teleoperate the robot, or recover the robot from failures. For each episode, the agent must choose between requesting human teleoperation, or using one of its autonomous controllers. In our approach, we learn to predict the success probability for each controller, given the initial state of an episode. This is used in a contextual multi-armed bandit algorithm to choose the controller for the episode. A controller is learnt online from demonstrations and reinforcement learning so that autonomous performance improves, and the system becomes less reliant on the teleoperator with more experience. We show that our approach to controller selection reduces the human cost to perform two simulated tasks and a single real-world task.