Is T-800 from Terminator not far away?
A couple of weeks back I saw this video on YouTube about this robot which is supposedly going to be “the best in it’s class” in this space. Being power by OpenAI automatically pushes it towards this claim but what can a company which makes LLMs do in the space of robotics. Then I saw the since then viral demo and well all I could say is, T-800 from Terminator might not be far away. Before I start fanboying about the movie that show “Skynet is the future of humanity”, let’s take a couple of steps back to see how we got here.
Remember Boston Dynamics, they came out with their robot dog Spot not so long ago and everyone in the tech community knew at that point things are getting serious. Of the countless reviews/hands-on documentations of Spot, a couple stand out to me. One was from MKDBH after which he stated
And then was Louis from Unbox Therapy who said “It is a moving intelligent….. I mean it is a Robot!” during the intro of their video.
At that point this was the most futuristic thing that anyone had seen. After 11 years of countless trials and errors they made a quad-pedal robot that could walk over terrains like never before.
After a year or so, we got to know about Tesla (the then most innovative company) having this idea of humanoid robots. This was Bumblebee. Tesla then came with future iterations of Optimus. A humanoid robot that is capable of performing repetitive tasks autonomously. We have seen a few woking demos. But the one where Optimus folds a shirt shows off it’s dexterity.
Features for Optimus Generation 2, as stated by Tesla are “Tesla-designed actuators and sensors, faster and more capable hands, faster walking, lower total weight, articulated neck, and more.” But one important thing to note here is all these movements are human operated i.e. not autonomous.
Then on 14th March 2024, Figure dropped a Demo of their brand new robot Figure 01 and this was maybe the first time the thought of Terminator came to my mind. With the backing of OpenAI, Figure had managed to create a fully autonomous robot that is capable of performing dexterous actions that it has learned. Keyword being learned. As per Figure, the actions shown on the demo were not controlled by any human wearing a controller or some sort of navigation gear but by itself. This boggled my mind to begin with. Combined with the Speech input and the stutter in between responses when asked with a complex question gave me an eerie feeling.
While combing though the demo, a couple of patterns I could notice:
- Each specific request put forward to Figure 01 took about 4–4.5 seconds for it to process and respond.
- Instead of a monotone voice, there was a subtle pause or in some cases a bit of a stutter in the response. This is not particularly new as there are Text-Speech LLMs out there where you can leverage these type of output but combined with voice and other functions, the whole package is unique.
There were complex questions asked where both motor functions and speech output were required. Over all, it took Figure 3 questions, “Hey Figure 01, what do you see right now?”, “Can I have something to eat?” & “Can you explain why you did, what you did, while you pick up this trash?”to make me think Skynet is near and there is nothing I can do about it.
For the sake of comparison, what would happen if I ask ChatGPT 3.5 the same questions after setting the context. Prompt might be a bit difficult to ensure that the scenarios are similar but it was worth a shot.
Setting the scene:
Without the need of asking the first question, ChatGPT 3.5 summarised the setting, but I asked it regardless.
In both cases, the summary with the limited amount of inputs make the most sense. Time to move on to a more complex one:
Similar to Figure 01, it gives the similar response. So far 2 for 2. Before asking the third question, there is the additional factor of introducing the garbage and basket. Let’s see what happens:
The response makes sense for just an LLM. It suggests a couple of steps to be taken to clean the garbage bags along with an one-liner for clean workspace. If a free-to-use LLM like ChatGPT 3.5 is capable enough to get so close to the desired output, imagine what the pioneers have in store. You might say “paranoid much!”, but my response after going through all these is “am I?”.