At present, autonomous driving is still closer. Robots are at least five years away. The long-term certainty is very strong. Bullish in the long term, bearish in the short term.
砺
Finally have time to reply to you. Hello, colleague.
I believe embodied intelligence is a scam, and the blame for this should primarily fall on academia rather than industry. This concept was initially hyped up by academia, but in industry, "robots" are not a new term. Robot learning itself is not a scam, but the claim that "general embodied intelligence is coming soon" is almost certainly a scam.
I graduated with a bachelor's degree from a second-tier university and am now pursuing a master's at Xinjiang University, with limited guidance from my supervisor. When reproducing results, I often fall into deep self-doubt because many papers at the Oral level suffer from the common problem of overly narrow contributions paired with overly broad titles and dissemination. High-level papers make strong engineering abstractions appear very concise, then package them as general embodied intelligence through titles, abstracts, and dissemination.
When it comes to industry, it gets even more interesting. The companies I've been in contact with (mainly small and medium-sized enterprises; I don't have access to large enterprises) that claim to work on embodied intelligence robots are essentially just adding an LLM to the old RTOS, performing some simple RAG or rule-based natural language instructions. Many home robot demos rely on human-in-the-loop remote control. Remember Figure AI? They also only use a small number of robots for simple tasks in factories. Regarding evaluation standards and simulation granularity, the current industry is a complete mess, generally lacking sufficiently fine-grained physical information, which is precisely the precision required for dexterous manipulation. And then there's Musk's Optimus; during its launch in 2024, the on-site interaction also heavily relied on manual operation.
It's a basic rule that industry leads academia in practice, and academia leads industry in theory. But now, even the strong industry hasn't produced epoch-making results, and even the strong academia hasn't produced truly epoch-making theories. So, using these results to seek financing, isn't that just hype?
You might mention VLA. VLA is not epoch-making. It essentially just stitches together existing advancements in "visual recognition, language understanding, and action generation" with an end-to-end network. The engineering integration is beautiful, but the underlying paradigm hasn't changed; it still lacks commonsense reasoning, physical intuition, and generalizable operational intelligence. But without changing this paradigm, it's hard to say there will be a breakthrough in the short term.
It's true that visual recognition is advancing, it's true that language planning is advancing, and it's true that imitation learning is advancing. But putting these things together is, at best, just a Plus version of RTOS, not even close to touching the concept of embodied intelligence.
