The future evolution of this field is a fascinating story.
We've seen the demos, and they're impressive. The real game-changer will be how we train these robots on the factory floor or at home.
Humanoids, robots mimicking biological systems, have been difficult to build due to the complex hardware of dozens of motors with 2-4 degrees of freedom. They're also expensive to maintain, especially if frequent calibration is needed.
Complex tasks require significant computing power, like that of a Jetson Orin. Users can't be expected to learn complex industrial programming languages.
Learning from example videos will likely be the easiest way to teach them new skills. Libraries of instructional videos, like how to sew a suit, will spread just like prompts do today.
Models can also be simplified. Text-to-motion models already exist as another use case.
The most interesting engineering challenge will be reducing costs to an acceptable level. Edge GPUs can offload traffic if they're close enough, and they can set micro movements with low latencies. Imagine smaller models describing how the robot should react to motion events until the next packet of model data arrives.
Complexity brings other engineering challenges. AI introduces more data blocks and code, like inference engines. Fail-safe operation is more critical than ever.
Reliability is key, and trace collection has always been the long-term solution. Smart models that can reproduce error cases will be invaluable. Systems need to rethink traditional UEFI reboot, restart, and reset cycles. Some information can't be lost; some information can't be updated frequently.
Engineering methods also need to improve. With more devices, we need robust ways to run robots reliably for extended periods. Models that can quickly converge to formal solutions to prove that an update is reliable are crucial for gaining market share.
Password-based security will also evolve. Frequent snapshots, monitoring, and easy restores can prevent data loss. Smartly differentiating large, publicly available data blocks from highly sensitive secrets can focus security efforts where they matter most.
Good dashboards and cost visibility can support purchasing decisions. Scaling up can improve AI user experience, especially latency issues in the short term. Long-term data processing issues can only be resolved by scaling out systems with more GPUs.
Some companies are already experimenting with universal data branching, like TiDB. This allows data, especially model data, to be forked and restored just like code versioning. Production and testing versions can be switched on the fly, enabling parallel exploration of multiple solutions.
Red teaming and security solutions may also evolve. Retrying logic to handle network errors (trying each call three times) can be exploited by attackers. The first failure reveals the target server, and the attacker can fail over to man-in-the-middle attack servers. Security issues may be masked this way or falsely point to infrastructure. Ensuring a robust backend can prevent extra debugging time and enable all sorts of edge use cases.
Modern databases are moving beyond write-ahead log streams to resolve network and CPU bandwidth issues in indexing. Cascading replication and streaming systems can make data processing more robust.