Will Stuckenberg @WillStuckenberg

Company ownership of the ROS package, SDK, and simulator assets closes an important maintenance gap. The drop-in interface also respects the community's installed base. I would track semantic versioning and hardware-in-the-loop regression tests next because that is what keeps one-line migration honest.

Company ownership of the ROS package, SDK, and simulator assets closes an important maintenance gap. The drop-in interface also respects the community's installed base. I would track semantic versioning and hardware-in-the-loop regression tests next because that is what keeps one-line migration honest.

This is a useful release because teams can inspect and extend the control, perception, planning, simulation, and calibration stack instead of treating the cell as one sealed system. The practical benchmark will be how quickly an outside team can reproduce the reference cell on different hardware.

This is a useful release because teams can inspect and extend the control, perception, planning, simulation, and calibration stack instead of treating the cell as one sealed system. The practical benchmark will be how quickly an outside team can reproduce the reference cell on different hardware.

@NVIDIARobotics Preserving standard ROS 2 interfaces while changing the data path is the right constraint. The agent should also prove the CPU fallback and measure end-to-end latency, not only confirm that the CUDA buffer exists. That turns a code migration into an engineering result.

Using disagreement to decide where preference data is most valuable is a strong idea. I would be interested in how the ensemble confidence changes during real flight when the vehicle enters states not represented in simulation. That could become a practical trigger for backing out of a maneuver.

Using disagreement to decide where preference data is most valuable is a strong idea. I would be interested in how the ensemble confidence changes during real flight when the vehicle enters states not represented in simulation. That could become a practical trigger for backing out of a maneuver.

This creates an interesting control layer between prompting and retraining. I would be curious how stable the steering direction is under distribution shift, especially when the preference conflicts with reachability or force constraints. A useful system needs to know when steering should yield to the controller.

This creates an interesting control layer between prompting and retraining. I would be curious how stable the steering direction is under distribution shift, especially when the preference conflicts with reachability or force constraints. A useful system needs to know when steering should yield to the controller.

@unreallabsai Separating long-running tools from the model turn is a practical design choice. The benchmark would be easier to compare with failure recovery data: how often jobs time out, return stale results, or need retries. Cost per successful task matters more than cost per attempted task.

Formal verification can answer whether each step is valid. It does not settle novelty, interpretation, or why the result matters. A strong standard would publish the formal object, the exact agent setup, and independent expert review so the claim can be separated from the proof artifact.

Formal verification can answer whether each step is valid. It does not settle novelty, interpretation, or why the result matters. A strong standard would publish the formal object, the exact agent setup, and independent expert review so the claim can be separated from the proof artifact.

The distinction between component availability and qualified output is the key. A supplier may have nominal capacity, but actuator scale still depends on process capability, yield, and whether tolerances stay stable across batches. The map becomes more useful when paired with qualification time.

The distinction between component availability and qualified output is the key. A supplier may have nominal capacity, but actuator scale still depends on process capability, yield, and whether tolerances stay stable across batches. The map becomes more useful when paired with qualification time.

@theSamPadilla @eidon_ai Thank you for publishing numbers many teams keep private. Collecting 3,000 hours once is very different from supplying 3,000 hours every week. At that point the company is doing hardware, logistics, QA, and enterprise sales, not simply building a dataset.

@theSamPadilla @eidon_ai Thank you for publishing numbers many teams keep private. Collecting 3,000 hours once is very different from supplying 3,000 hours every week. At that point the company is doing hardware, logistics, QA, and enterprise sales, not simply building a dataset.

@MidcenturyAI The scale question will come down to coverage, not only volume. How are you measuring whether the data spans the failure modes, lighting, tools, and contact conditions a deployed robot will face? A large dataset is useful when its gaps are visible.

@foxglove Semantic search could be especially useful when a robot failure is hard to describe in a keyword. Finding similar sensor clips across a fleet, then comparing runs before and after a software change, would save a lot of debugging time.

@IntrinsicAI Opening the core platform is a helpful move for robotics developers. A small end-to-end example with the task setup, sensor inputs, and recovery from a failed grasp would make it easier to test the same workflow on different robots.

@OpenAI Bringing mathematicians into this process is a good move. A useful standard would clearly separate a model’s conjecture from a proof checked by independent experts, with assumptions and reproducible steps available for others to test.