In the past, when evaluating robots, we often started by looking at how many actions they could perform, how smoothly they moved, and how powerful their hardware was. But I believe these are just surface-level indicators. What will truly determine a robot’s potential for future development is whether it can keep improving through repeated task execution.
That’s because the real world is far more complex than a laboratory. A seemingly simple grasping task involves multiple steps, including object recognition, position estimation, motion planning, and force control. An error at any one of these steps can cause the task to fail.
If a robot has an 80% task success rate, then over 100 independent attempts, it would fail an average of 20 times. This figure is just a hypothetical example to illustrate the point; it is not a measured result from Axis.
What’s really worth paying attention to is whether those 20 failures can help the system identify problems. After adjusting its strategy, can it improve its success rate?
That’s what makes Physical AI so interesting.
One important reason I’m following @axisrobotics is its training and feedback loop:
Human demonstration → strategy generation → robot execution → outcome feedback → strategy optimization → execution again.
It may look like a simple loop, but it determines whether a robot can move beyond one-time training and improve continuously through iteration.
For example, if the same grasping task needs to be practiced 100 times, what really matters isn’t simply completing those 100 attempts, but whether each attempt provides useful information. Which actions are more reliable? Which environments are more likely to cause failures? What strategies can adapt to different objects?
When these questions can be analyzed using data, training has the potential to shift from simply increasing the number of attempts to targeted optimization.
Of course, increasing the number of training attempts doesn’t necessarily lead to linear gains in capability. Data quality, feedback accuracy, and the ability of strategies to generalize are all important.
I’m more excited to see Axis continue making progress in these areas, so that robots don’t just follow fixed programs to perform actions, but can gradually improve their performance through effective feedback.
Hardware determines what a robot can do, while its training mechanisms determine how well it might be able to do it. If more efficient training and feedback can reduce the cost of repeated trial and error while improving robots’ reliability across different environments, then the potential applications of Physical AI could expand even further.
Rather than focusing only on how impressive a single demonstration is, I’m more interested in whether the system can build up experience over time and truly turn that experience into capability.
So, I’m staying optimistic!
#AxisRobotics