Winding up 2024
Best wishes for 2025
Merry Christmas to all those who celebrate it! This will be the last post of the year. Learn With A Robot now has a great community (Many thanks to all of you for participating), having doubled the number of newsletter subscribers this year, and over 500 subscribers in our YouTube channel (Please subscribe to the YouTube channel for latest videos). With greater participation in the community, we get unique insights into what is interesting to the audience. I wanted to use this year ending post to summarize what has been hot in robotics and what will likely trend in 2025.
Trends
Robotics and AI are intertwined. Not long ago, robotics and Artificial Intelligence were separate areas of study and research. The progress in robotics was measured by research in the mechanical aspects of robotics, such as sensors and actuators. As an example, if we look at the effort spent by Anki Inc. in building the Vector robot (2013-2018 time frame), most of it went in designing the form factor that could fit the sensors and motors, and in building the animations for an emotional robot. The use of Artificial Intelligence was mainly on path planning (SLAM). Fast forward to 2024, the main topic in robotics is “When will robotics meet its ChatGPT moment”. In the Humanoids Summit last week, many speakers raised this question. While some thought that this was inevitable in 5 years, many agreed with Jonathan Hurst, the co-founder of Agility Robotics (the makers of Digit),who said that robotics engineers would need to earn the Chat GPT moment, and that robotics is still away from its ChatGPT moment because of the much higher degrees of freedom that need to be solved to make robots successful. Regardless of the time it takes to get there, there is widespread understanding that AI will expose robotics to the masses by delivering useful services. The high evaluations in emerging startups like Skild.ai and Physical Intelligence reflect this viewpoint.
Reinforcement Learning: We will see robots performing unimaginable moves with the help of Reinforcement Learning (RL). To get a feeling of what RL can achieve, take a look at this recent video from Unitree. The video shows a quadruped on wheels make effortless moves like backflips, going down a rugged terrain, wading across water, jumping over obstacles, jumping from a building and landing safely, and many other unbelievable moves. In case you question it, this video is real, not generated by AI. With a correctly defined reward function, RL can accomplish amazing tasks. I am spending some time in the holidays learning how to apply RL to the Petoi Bittle robot, and I sense immense possibilities in this area.
Large Language Models (LLMs) will dominate the news. The progress in LLMs resembles something like attempting to climb a mountain range. You need to find the next hill to climb and reach the next summit. From there on, you scan the horizon and identify the next summit to surmount. 2024 has seen major progress in reasoning models with OpenAI’s o1 and recently o3 dominating the news, and quickly caught up by open source reasoning models such as QwQ. Not to be left behind, Google released its reasoning model called “Gemini 2.0 Flash Thinking Experimental” for initial evaluation. Vision Language Action (VLA) models which are more applicable in the field of robotics, have made a lot of progress this year (See our related article on OpenVLA). 2025 is sure to see rapid progress in VLA models.
Humanoids. Humanoids made major advances in 2024. Believe it or not, you can now rent a humanoid to show up at your party and greet guests. The humanoid is a great crowd puller. As an example, take a look at our video of Unitree G1 robot at the Humanoids Summit.
Many more commercial releases of humanoids are coming up in 2025. As per Andra Keay, there is already a lineup of 40+ humanoids slated for commercial availability in 2025. That’s a lot to choose from. While many used to question the humanoid form factor, we saw a broad acceptance of humanoid based robotics solutions in 2024. There is a broad understanding and consensus that humanoid robots have advantages of operating in regions where humans need robots most. These include staircases, apartment complexes, and small places (where robots need to crawl into). In 2025, we might see the launch of the Tesla Optimus robot. We are also likely to see several deployments of the Figure robot beyond what they are doing for BMW.
Wind up
As we draw the curtains for 2024, I would like to thank you all for reading these notes and joining the journey of learning with robots. Warm wishes for the holidays and 2025. See you in the new year!