Search for: All records

Object insertion under tight tolerances (less than 1mm) is an important but challenging assembly task as even small errors can result in undesirable contacts. Recent efforts focused on Reinforcement Learning (RL), which often depends on careful definition of dense reward functions. This work proposes an effective strategy for such tasks that integrates traditional model-based control with RL to achieve improved insertion accuracy. The policy is trained exclusively in simulation and is zero-shot transferred to the real system. It employs a potential field-based controller to acquire a model-based policy for inserting a plug into a socket given full observability in simulation. This policy is then integrated with residual RL, which is trained in simulation given only a sparse, goal-reaching reward. A curriculum scheme over observation noise and action magnitude is used for training the residual RL policy. Both policy components use as input the SE(3) poses of both the plug and the socket and return the plug's SE(3) pose transform, which is executed by a robotic arm using a controller. The integrated policy is deployed on the real system without further training or fine-tuning, given a visual SE(3) object tracker. The proposed solution and alternatives are evaluated across a variety of objects and conditions in simulation and reality. The proposed approach outperforms recent RL-based methods in this domain and prior efforts with hybrid policies. Ablations highlight the impact of each component of the approach. 

Object insertion under tight tolerances (less than 1mm) is an important but challenging assembly task as even small errors can result in undesirable contacts. Recent efforts focused on Reinforcement Learning (RL), which often depends on careful definition of dense reward functions. This work proposes an effective strategy for such tasks that integrates traditional model-based control with RL to achieve improved insertion accuracy. The policy is trained exclusively in simulation and is zero-shot transferred to the real system. It employs a potential field-based controller to acquire a model-based policy for inserting a plug into a socket given full observability in simulation. This policy is then integrated with residual RL, which is trained in simulation given only a sparse, goal-reaching reward. A curriculum scheme over observation noise and action magnitude is used for training the residual RL policy. Both policy components use as input the SE(3) poses of both the plug and the socket and return the plug's SE(3) pose transform, which is executed by a robotic arm using a controller. The integrated policy is deployed on the real system without further training or fine-tuning, given a visual SE(3) object tracker. The proposed solution and alternatives are evaluated across a variety of objects and conditions in simulation and reality. The proposed approach outperforms recent RL-based methods in this domain and prior efforts with hybrid policies. Ablations highlight the impact of each component of the approach. 

Every mammal studied to date has been found to have a male mutation bias: male parents transmit more de novo mutations to offspring than female parents, contributing increasingly more mutations with age. Although male-biased mutation has been studied for more than 75 years, its causes are still debated. One obstacle to understanding this pattern is its near universality—without variation in mutation bias, it is difficult to find an underlying cause. Here, we present new data on multiple pedigrees from two primate species: aye-ayes (Daubentonia madagascariensis), a member of the strepsirrhine primates, and olive baboons (Papio anubis). In stark contrast to the pattern found across mammals, we find a much larger effect of maternal age than paternal age on mutation rates in the aye-aye. In addition, older aye-aye mothers transmit substantially more mutations than older fathers. We carry out both computational and experimental validation of our results, contrasting them with results from baboons and other primates using the same methodologies. Further, we analyze a set of DNA repair and replication genes to identify candidate mutations that may be responsible for the change in mutation bias observed in aye-ayes. Our results demonstrate that mutation bias is not an immutable trait, but rather one that can evolve between closely related species. Further work on aye-ayes (and possibly other lemuriform primates) should help to explain the molecular basis for sex-biased mutation. 

We present a physical neural network (PNN) approach towards multiplane light conversion (MPLC) design. PNN performs a full parameter search with flexible optimization pathways and can tune various design attributes as hyperparameters.