The Robot Report Podcast · 16 Apr 2021 · 1h 17m
Ken Goldberg and Jim Liefer of Ambi Robotics discuss their Sim2Real AI approach to robotic grasping, explaining how they trained neural networks on simulation data to handle diverse e-commerce objects without human intervention. Paul Mikesell, founder of Carbon Robotics, details the development of their third-generation autonomous weeding robot, which uses high-powered lasers and computer vision to identify and destroy weeds in row crops. The episode also covers WTWH Media's acquisition of Mobile Robot Guide and updates on the Mars 2020 Ingenuity helicopter.
Where this conversation describes AI being used to build or deliver a product. Part of the AI in practice tracker.
Engineering & simulationDexnet / Sim2Real AIretrain time from weeks/months to overnight
Ambi Robotics uses a Sim2Real AI platform to train neural networks in simulation by injecting specific noise to match real-world sensors, allowing them to retrain the grasping model overnight rather than taking weeks or months of physical testing.
“if you change the gripper, we rerun the simulation overnight and retrain a new neural network to run the next day.”
Ken Goldberg · Hear it at 29:27
Engineering & simulationDeep learning neural networks
Carbon Robotics trains deep learning neural networks using labeled imagery to distinguish weeds from crops, requiring continuous data collection and retraining as soil and crop conditions change throughout the season.
“We gather imagery, and then we have a labeling system … we run back through the through the neural net and train it. … we will continually gather more images to fine tune that neural net.”
Paul Mikesell · Hear it at 54:28
early-stage companies
Other shows on the same ground, closest first.
Hosts Daniel and Farbod discuss how computer vision AI is applied in manufacturing and agriculture, featuring insights from Plain Sight AI and a project by the Technical University of Denmark. They explain Plain Sight AI's three levels of vision systems for detecting hazardous objects on production lines, emphasizing that Level 3 instance segmentation offers superior sorting and counting capabilities. The hosts also detail a robot developed by PhD student Ronja Goldring to autonomously remove weeds from organic dairy farms using laser or electrocution, highlighting the importance of large, diverse training datasets and collaboration with end-users.
Dr Petra Bosilj of the University of Lincoln explains that robotic vision in agriculture is bottlenecked by a lack of annotated training data for complex tasks like distinguishing weeds from crops, though precision spraying can reduce herbicide use by 80%. Chris Chavasse, founder of Muddy Machines, describes his startup's development of lightweight, swarm-capable field robots for harvesting tender stem broccoli, designed to reduce soil compaction compared to heavy industrial machinery. The episode contrasts academic research into phenotyping and soil health with the commercial challenges of building rugged, soft-robotic end effectors and securing funding during the pandemic.
Vassia Simaiaki, Head of AI Vision and Graphics at Wayve, explains the company's end-to-end AI approach to autonomous driving, which relies on camera data rather than expensive hardware like LiDAR or HD maps. She details how Wayve uses novel view synthesis and a neural simulator called NeuroSim to generate synthetic data, allowing their models to generalize to unseen scenarios like the 'plastic bag problem' or complex London traffic. The discussion highlights Wayve's partnership with Asda to expand grocery delivery coverage and their focus on proving safety through simulation to build public trust.
Edward Johns, Director of the Robot Learning Lab at Imperial College London, discusses the challenges of robotic manipulation, arguing that the bottleneck is currently AI and control rather than hardware. He explains his lab's use of imitation learning, where robots learn by being physically guided, and details the Dali Bot project, which uses AI image generators to help robots visualize task goals. Johns outlines applications ranging from warehouse sorting to flexible manufacturing and long-term domestic robots, while noting the significant data scarcity in robotics compared to other AI fields.
Michelle Lu, co-founder and CEO of Vsim Technology, explains how her company uses GPU-accelerated physics simulation to accelerate robot learning and close the sim-to-real gap. Previously at NVIDIA creating Isaac Gym, Lu details how Vsim overcomes the scalability and performance limitations of traditional CPU-based reinforcement learning frameworks. The discussion highlights specific challenges in simulating tactile sensors, deformable objects, and tendon systems for the ARIA Robot Dexterity program, aiming to enable complex dexterous manipulation tasks.
The summary, chapters and topics are generated from the episode’s audio by this directory, not written by The Robot Report Podcast. The episode itself belongs to its publisher.