Daniela Rus: Physical AI

The Vision of Physical AI

• 💡 Physical AI integrates advanced AI capabilities like understanding text and images with physical machines to enhance their intelligence and decision-making.
• 🤖 Current challenges include AI being confined to computers, robots lacking inherent intelligence, and AI models making mistakes in safety-critical applications.
• ⚡ Existing frontier AI models (e.g., ChatGPT, Claude) have high energy costs due to large parameter counts and energy-intensive pre-training and fine-tuning phases.
• ⚠️ These large models also struggle with diminishing returns in performance as they scale, and their size prevents deployment on small, edge devices.

Advancing Energy Efficiency

• 🚀 Techniques like pruning (removing unimportant parameters) and quantization (reducing weight precision) are used to simplify models without losing effectiveness.
• 📈 DeepSeek demonstrated a 45-fold increase in training efficiency for large AI models, achieving high performance with significantly reduced cost and energy.
• 🔋 Despite these advancements, the fundamental architecture of current large models still presents challenges for energy-efficient deployment on physical machines.

Introducing Liquid Networks

• 🧠 Inspired by the simple nervous system of the C. elegans worm, Liquid Networks offer a new class of AI algorithms with fewer, more complex neurons.
• ✅ These networks use differential equations and are designed to adapt after training, making them provably causal and highly energy-efficient.
• 🚗 In self-driving car examples, Liquid Networks use significantly fewer neurons (19 vs. 100,000) and exhibit cleaner, more focused attention maps than traditional deep networks.
• 🌳 They demonstrate strong out-of-distribution adaptation, performing consistently even when environmental contexts change (e.g., seasonal variations in driving).

Enhancing Safety and Capabilities

• 🛡️ Barriernet integrates control barrier functions, a mathematical layer, with machine learning models to ensure provably safe outputs for AI in critical applications.
• 🎯 This approach guarantees that the AI's actions remain within a defined safe region, preventing errors in tasks like autonomous driving, flying, and robot swarm navigation.
• 🌐 Liquid Foundation Models, developed by Liquid AI, show higher performance on benchmarks like MMLU Pro and have a smaller memory footprint, making them suitable for edge devices like phones and robots.
• 🗣️ These models can also handle multimodal data and integrate language to elevate reasoning from raw sensor data to higher-level concepts, improving adaptation to new scenarios.

AI for Physical Creation and Learning

• 🛠️ AI can be used for physical design, enabling "text-to-robot" and "image-to-robot" capabilities by generating designs that adhere to the laws of physics and task requirements.
• 🔄 A diffusion process combined with differentiable simulation refines these designs, accelerating the innovation cycle for new physical products.
• 🧑‍🤝‍🤖 Robots can learn complex tasks from humans through human-to-robot learning, by collecting sensor data (pose, muscle, gaze) from human demonstrations to replicate delicate movements.
• 🌱 This allows for machines that move with grace and precision, performing tasks like food preparation and cleaning in a more human-like manner.

The Future of Physical Intelligence

• 🌟 The combination of liquid networks, physical design, and human-robot learning promises extraordinary benefits, including personal assistants and bespoke machines.
• 🌍 This technology can extend human reach, refine precision, and amplify strength, opening new ways to interact with the physical world.
• ⚖️ However, the development of physical intelligence must proceed with a guiding hand, emphasizing responsibility for the planet and all living things.
• 🔍 Further research is needed into provably causal solutions and the integration of diverse sensory modalities (e.g., lidar, smell sensors) to enhance robot capabilities and safety.