Mankaran Singh, the founder of Flowdrive and an autonomous vehicles engineer at Ola Electric, delivered an insightful talk at the Machine Learning Developers Summit (MLDS) 2024 in Bengaluru. The summit, India’s largest generative AI conference, showcased the latest advancements and perspectives in the field. Singh’s presentation focused on the current state of self-driving cars and robotics, shedding light on the industry’s challenges and potential solutions.
Mankaran Singh, a prominent figure in autonomous vehicle engineering, took the stage at MLDS 2024 to share his journey and groundbreaking work at Flowdrive. As the founder, Singh initiated the Flowdrive project two years ago with a small team of three enthusiasts. Since then, it has evolved into a thriving community of a thousand individuals. His talk delved into the transition from training foundation models for self-driving cars to the ambitious venture of training models for general-purpose robotics.
The Three Approaches
Singh started by outlining the historical approaches the industry has taken in developing self-driving cars. The classical approach, rooted in basic mathematical formulations without involving machine learning, proved insufficient. The second approach, utilizing machine learning for tasks like object detection and lane recognition, faced challenges despite substantial investments. The third and most promising approach, according to Singh, is the end-to-end model, exemplified by companies like Tesla and comma AI.
Lessons from Robotics
Drawing parallels between robotics and self-driving cars, Singh highlighted the success of end-to-end models in solving complex problems. He emphasized the limitations of relying on intermediate steps, such as object detection, by citing examples from the robotics industry. Companies that have shifted to end-to-end approaches, like Tesla, have shown significant progress, underlining the importance of this paradigm shift.
Data Challenges and Industry Shift
Addressing the audience, Singh pointed out the critical challenges in the industry, including the need for massive datasets and the high costs associated with training models. He disclosed that companies embracing end-to-end approaches, like Tesla, have already shifted their strategies. The talk featured insights into the iterative process of training models, emphasizing the challenges and costs involved.
Beyond Self-Driving Cars
Singh expanded the discussion to encompass broader applications of end-to-end models, particularly in general-purpose robotics. He touched upon the need for fine-grained control in tasks involving human hands, citing examples like packaging gifts and handling delicate objects. The hardware capabilities, illustrated by projects like Aloha’s cable-zipping robot, showcased the potential for real-world applications.
Challenges Ahead
While optimistic about the future of robotics, Singh acknowledged the challenges ahead. Collecting labeled data for robotics poses a unique obstacle, unlike the wealth of labeled data available for vision models on the internet. Safety concerns were also addressed, as deploying models into physical bodies demands meticulous validation to avoid potential harm.
Conclusion
Singh concluded by outlining the ambitious goals of his initiative, Tensor Hard AI. With a mission to release general-purpose robotic solutions within two years, the project aims to bridge the gap between hardware capabilities and AI software. He invited the audience to follow his journey on Twitter, providing a glimpse into the vision for the future of robotics.
Mankaran Singh’s talk at MLDS 2024 offered a comprehensive view of the evolving landscape of self-driving cars and robotics, emphasizing the pivotal role of end-to-end models in shaping the future of mobility and automation.