Foxglove

This week in Physical AI: Gemini 4 Argon taught Apollo cornhole from a video, Pi opened its API, Rhoda and NEURA added touch, NVIDIA and Artificial Analysis priced out LLMs as robot policies, Nature Will shipped a 50B hand model, Waymo opened three cities, Standard Bots added prompting, and Agility sped up Digit 5 charging. Google DeepMind Gemini 4 Argon built a MuJoCo sim from a human cornhole video and used it to get Apptronik's Apollo to sink 7 throws in a row. From one tabletop photo it also generated 4,500+ sim episodes that took a real VLA from 56% to 75% success. https://lnkd.in/gttgMVEy Physical Intelligence opened the Pi API beta, so teams can fine-tune π on their own demos and run it through hosted inference. They flag network connectivity as a pain point for some cloud deployments. https://lnkd.in/gS-Q4bEA Rhoda AI built a thin fingertip sensor (a few mN to 40+ N, plus shear) and put it on a handheld UMI so demos capture force. With its cameras blocked, the robot found textured balls by touch, including ones it never trained on. NEURA Robotics's Neuron AI also folds touch into one model across five robot bodies. https://lnkd.in/gN9y3kSp https://lnkd.in/gcViHYaS NVIDIA Robotics's Astranomy ran GPT-6 Astra as a policy on RoboLab-120 in sim and hit 74.8% versus 28% for π0.5, at 13 to 18 s per call with the sim paused and ~$36 per success. On a real arm, Artificial Analysis found models often reached a passing state and then kept adjusting until they broke it. About 70% of their cost was input context, driven by all the camera imagery. https://lnkd.in/guBxjb2T https://lnkd.in/gZmCdunH Nature Will's Natura-1 is a 50B MoE pretrained on 1M+ hours of egocentric human data, driving dexterous hands zero-shot from language on a 100 ms loop. Company claims plus a demo for now. https://www.naturewill.ai Waymo opened driverless rides to anyone in San Antonio, San Diego and Tampa, including San Antonio's airport. https://lnkd.in/gstZCpg9 Standard Bots lets operators prompt for a routine or tap a part in the camera feed instead of teaching waypoints, which they say cuts programming time 90%. https://lnkd.in/ga5HxaNZ Agility Robotics's Digit 5 will run 90 minutes per 9-minute charge with Nyobolt's battery, up from 2:1 on Digit 4, for 20+ hours of work a day. https://lnkd.in/gKhyfHEe

One more thing we quietly shipped this week: Basic seats. You can now invite unlimited users to Foxglove, with read-only access to existing layouts and cloud data, at no additional cost. This is a big deal. Many people told us that they wanted to share links with their broader team (product, QA, executives, etc) - but they couldn't justify paying full price for those users. Basic seats lower the barrier to entry, making it easier to share Foxglove across your organization. This on top of our new $20/mo Pro plan makes it easier than ever to get started with Foxglove!

Internal tooling can work well for a small team. But as engineers, robots, and data volumes multiply, maintaining it becomes its own job. Scale your debugging workflow with Foxglove.

We founded Foxglove with a two part thesis. At the time, both parts were highly non-consensus: 1. There are going to be billions of autonomous machines deployed over the coming decade. If we already have the technology to make cars drive themselves around San Francisco, imagine everything else humans do today that can be automated. The technology is here, just not widely distributed. 2. Infrastructure is the biggest bottleneck. The first generation of AV companies had to build all their infrastructure in-house with thousands of engineers - data infra, sim infra, training infra, fleet infra. The next generation of physical AI companies cannot succeed without great off-the-shelf infrastructure. Of these, #1 is now a pretty mainstream belief. But #2 still feels pretty non-consensus. This feels good, if i'm not building against the mainstream consensus I tend to get uncomfortable.

Most robotics companies think they have an autonomy problem. They actually have a data problem 👉💾 Models are only as good as the data they are trained on. But collecting high-quality robotics data is expensive. Annotating it is slow. Finding the edge cases that matter is like searching for needles in a terabyte-scale haystack. The companies that build effective data flywheels will outpace everyone else: 1. Collect comprehensively, not selectively 2. Search and curate by behavior, not just timestamp 3. Close the loop from deployment failures back to training Simulation helps, but real-world data is irreplaceable for validation. The question is whether you can find and use that data fast enough to matter. Every robotics company eventually becomes a data company. The ones that realize this early invest in the infrastructure to support it.

Another big launch: Bring Your Own Storage (BYOS)! 🎁 BYOS lets you connect an existing storage bucket (e.g. Amazon S3) to Foxglove. Paired with our new Data Search feature, this means you can now index, search, and visualize data without needing to deploy or manage any additional infrastructure. Read more: https://lnkd.in/ez22Hsny