𖥂🎮Embodied AI - Challenges

In recent years, LLMs have made AI center of attraction for once more. There are cognitive tasks which are now feasible due to LLMs but where is my robot that could sweep my floor, vacuum carpets, water the plants and likely other physical task in and around the house?

Embodied AI, has its own unique challenges. There is not enough data for physical AI to train on when compared to what we have for LLMs

Biological designs are optimized for functioning in physical world, Studying bio designs and implementing the findings to robot designs needs a lot of work

Robots can generate commands using LLMs but the code snippets to achieve these are pre-generated. However, for local awareness, robot must be capable for generating these movement and sensing commands on the fly to perform a complex task. It must also learn the best combination for doing it, so training on the job is must.

Learning from physical world is a huge limitation. The simulation infrastructure can help the ecosystem a lot.

I would imagine a RoboLLM for autonomous robot command and operations in near term same as LLMs like Claude, Gemini Llama are for generating text.

The day when autonomous drone will surpass skills of human pilots is not too distant. Like AI mastered chess and Alpha go in last couple of decades, this is next frontier in the third decade of 21st century