Intelligent Robots in 2026: Are We There Yet? with Nikita Rudin - #760 - TWIML AI Podcast Recap

Podcast: TWIML AI Podcast

Published: 2026-01-08

Duration: 1 hr 7 min

Guests: Nikita Rudin

Summary

Nikita Rudin, CEO of Flexion Robotics, discusses the current state and future of robotic autonomy, emphasizing the challenges in closing the sim-to-real gap and the role of reinforcement learning. He highlights a modular approach to robotics as more effective than end-to-end models.

What Happened

Nikita Rudin, co-founder and CEO of Flexion Robotics, elaborates on the current challenges and advancements in deploying fully autonomous robots. He highlights the impact of reinforcement learning and simulation on robotic locomotion, demonstrating how these technologies can significantly reduce training time for quadruped robots from weeks to minutes using GPUs and parallel simulators.

Rudin delves into the complexities of the sim-to-real gap, particularly when visual inputs are introduced, which adds noise and complicates the transfer from simulation to real-world applications. He notes that blind locomotion, devoid of perception inputs, remains more robust and easier to train.

The conversation explores the debate between end-to-end models and modular approaches, with Rudin advocating for the latter. He argues that separating locomotion, planning, and semantics is practical, given the current technological constraints and the need for more refined simulations.

Rudin introduces the concept of 'real-to-sim,' where real-world data refines simulation parameters for better training fidelity. This approach, combined with reinforcement learning, imitation learning, and teleoperation data, helps train robust policies for both quadruped and humanoid robots.

He also discusses Flexion Robotics' hierarchical approach to robot training, utilizing pre-trained Vision-Language Models (VLMs) for high-level task orchestration alongside Vision-Language-Action (VLA) models and low-level whole-body trackers. This framework allows for effective task orchestration despite the challenges in physical interaction.

The episode highlights the ongoing development in humanoid robots, which are still not fully autonomous or ready for home use but show promise in industrial applications. Rudin emphasizes that humanoid robots are in a beta phase, with companies like 1X testing the waters for consumer deployment.

Rudin reflects on the challenges of reinforcement learning in real-world scenarios, noting the expense and difficulty compared to simulations. He points out the complex nature of human reward signals and the lack of tactile feedback in robots, which complicates real-world training.

Finally, Rudin shares his outlook on the future of robotics, predicting that humanoid robots could start providing value by the end of next year, particularly in industrial settings. He suggests that sending humanoid robots to Mars could be a strategic move before human colonization.

Key Insights

• Reinforcement learning and simulation technologies have reduced the training time for quadruped robots from weeks to minutes by using GPUs and parallel simulators.
• Blind locomotion, which operates without perception inputs, is more robust and easier to train compared to systems that rely on visual inputs due to the complexities introduced by the sim-to-real gap.
• The 'real-to-sim' approach uses real-world data to refine simulation parameters, enhancing the training fidelity of robots through a combination of reinforcement learning, imitation learning, and teleoperation data.
• Humanoid robots are currently in a beta phase, with potential for industrial use by the end of the next year, and there is a strategic consideration for deploying them to Mars before human colonization.