Many users are excited about humanoid robots like Kinetiq Ascend doubling throughput with real-world reinforcement learning because it allows them to learn from real failures instead of copying humans.
Based on 2 visible X reactions from 6 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@Scobleizer @TheHumanoid This is exactly the breakthrough physical AI needed!!!!!!
@Scobleizer @TheHumanoidAI It’s awesome Robots learning from real failures instead of copying us
Most robots today learn by copying human demonstrations. That gets you far, but a robot that only imitates can never move faster or more reliably than the person it copied, and it never learns the cost of failure. What @TheHumanoidAI is doing with Kinetiq Ascend is the shift I have been waiting to see in physical AI. Instead of copying demos, their humanoids practice real production tasks and learn from their own successes and failures, with reinforcement learning running directly on real hardware, around the clock. The results are the part that stops me. On a picking and handover task, success went from 80 to 98 percent...a tenfold drop in failures. On bimanual tote handling, throughput more than doubled and success climbed to nearly 99 percent. All from days of robot time, not months. The bigger story is the curve. They are seeing the same predictable, compute driven improvement that reshaped language models, now showing up in robots that manipulate real objects on real production lines. And once these fleets ship, they keep learning on the job, every deployed robot becomes a source of training data. This is what the road to reliable, general purpose humanoids actually looks like. Not a demo, a system that gets better the more it works. Check it out here: https://thehumanoid.ai/technology/kinetiq-ascend/
Most robots today learn by copying human demonstrations. That gets you far, but a robot that only imitates can never move faster or more reliably than the person it copied, and it never learns the cost of failure. What @thehumanoid is doing with Kinetiq Ascend is the shift I have been waiting to see in physical AI. Instead of copying demos, their humanoids practice real production tasks and learn from their own successes and failures, with reinforcement learning running directly on real hardware, around the clock. The results are the part that stops me. On a picking and handover task, success went from 80 to 98 percent...a tenfold drop in failures. On bimanual tote handling, throughput more than doubled and success climbed to nearly 99 percent. All from days of robot time, not months. The bigger story is the curve. They are seeing the same predictable, compute driven improvement that reshaped language models, now showing up in robots that manipulate real objects on real production lines. And once these fleets ship, they keep learning on the job, every deployed robot becomes a source of training data. This is what the road to reliable, general purpose humanoids actually looks like. Not a demo, a system that gets better the more it works. Check it out here: https://thehumanoid.ai/technology/kinetiq-ascend/
Many users are excited about humanoid robots like Kinetiq Ascend doubling throughput with real-world reinforcement learning because it allows them to learn from real failures instead of copying humans.
Based on 2 visible X reactions from 6 accounts; directional sample.
Ask a question below.
Published answers will appear here.