Figure 01 robot demos its OpenAI integration
A new demo of Figure’s humanoid robot tightly integrated with OpenAI models showcases natural-language commands translated into fluid object manipulation, sparking both awe and skepticism. Commenters highlight the impressive text-to-servo control, vision-guided dexterity and human-like speech, while questioning how much is cherry‑picked, how general the capabilities really are, and whether latency and hardware limits will hold back real-world use. Many see this as an early glimpse of robots performing manual labor via conversational instructions, with potentially profound implications for automation and the relative value of physical versus knowledge work.
Overall impressions
- Many find the demo visually and conceptually impressive, especially seeing GPT-style capabilities embodied in a humanoid robot.
- Others are underwhelmed on the AI side, viewing it mostly as “GPT-4 + vision + function calling” hooked to a robot, with a very simple world model.
Text-to-motion and dexterity
- The speech-to-servo / text-to-action link is widely highlighted as the standout: the robot follows natural language commands to perform physical tasks.
- Commenters praise the speed and fluidity of object manipulation, noting that other robot demos often need sped-up video.
- Some emphasize how hard tasks like gently gripping an apple, placing plates into a drying rack, and passing objects to a moving hand are.
Demo realism, cherry-picking, and generality
- Many suspect heavy cherry-picking: numerous takes until a flawless run, idealized environment, and tasks tuned to this exact table, objects, and human position.
- Debate over whether movements are “pre-baked” vs. end-to-end neural policies triggered by the LLM; a linked high-level system description suggests neural controllers, but generality remains unclear.
- Several request demos in varied or cluttered settings, with arbitrary objects, to test robustness.
- There is explicit comparison to past over-produced AI/robotics demos (e.g., Bard, Gemini, stylized Boston Dynamics videos), feeding skepticism.
- One commenter outright claims the robot looks rendered/CGI; others do not substantiate this.
Voice, stutters, and latency
- Multiple people notice realistic “uh”s and stutters. Some find them creepy or “dumbed down,” others feel they increase naturalness.
- Several attribute this to current TTS systems (OpenAI, ElevenLabs-like), where lower “stability” or artifacts produce human-like imperfections.
- Latency is a recurring complaint; people argue that truly conversational robots require much faster end-to-end pipelines (ASR → LLM → control → TTS).
- Discussion touches on specialized low-latency inference hardware and multimodal models as critical bottlenecks.
Reasoning, behavior, and limitations
- The robot’s inferred behavior (e.g., throwing trash in the bin without being told) is praised as an example of LLM-style “common sense” / abductive reasoning.
- Others point out logical flaws, such as putting a presumably dirty plate directly into a drying rack; lack of memory is suggested as a cause.
- Some note that the “world” is still extremely constrained; general symbolic planning plus multiple neural subsystems is seen as the likely long-term path.
Labor, economics, and social impact
- Several see this as a big step toward robot manual labor, possibly preceding full displacement of many knowledge-work tasks.
- Others express deep pessimism that such tech will become broadly useful, affordable, or socially beneficial, expecting it to end up as marketing spectacle or military tech.
- There is debate over whether we’re finally overcoming the traditional “manual labor is harder than cognition” view, with some suggesting that once a capable robot exists, teaching it physical chores may be easier than high-level intellectual work.