@ErenChenAI I've heard that at an exhibition last year Engine AI refused to do a demo on stage because the friction of the floor was too different to what it's model was trained on. Looks like something similar here?
@teortaxesTex@SchmidhuberAI@dwarkesh_sp Statistically-driven simulators that provide real world feedback will likely be the way forward for online post-training, at least for robot policies
It looks like Zhipu AI is also doing a bit of VLA research. Don't be surprised to see more LLM labs in China follow a similar direction. VLA tends to generally pace alongside LLM research, with a DiT slapped on the end for action sampling and generalised action representation
@tuhin@NoriRobotics@weaverobotics@sundayrobotics At this point in time everyone should be more asking about how they're gonna manage on-policy data collection for when homes are out of distribution for the model than discussions about climbing stairs
A couple of questions I have about this being shipped after going through the website.
1. If a customer bought the robot and the customer's home is too outside the distribution of the training dataset, how exactly does teleoperation services work? Does the customer have to pay for teleoperation fees on top of a model they already paid for? How long does the customer have to wait for a person to come collect data?
2. The privacy policy on the website mentions that data collected will be shared with third party advertisement companies. Isn't that a little bit of an overstep in privacy? It makes sense to only keep data for post-training models
@tiagosada Folding clothes is just like the benchmark task in robotics that VLAs can reach 100% success rate which has some real-world use case. In China it's treated like a standard benchmark saying you've reached this level of complexity for robotics startups since it's quite difficult
Shipping robots and complementary models at this stage is more about pure grit and brute force than anything else right now. Until that magic path to physical intelligence is found the path will be paved with lots of pain and huge efforts of labour