Rerun

Rerun nails a specific pain point: teams working on robot learning often cobble together separate tools for data logging, visualization, querying, and training. Rerun's SDK replaces that stack with one consistent API, which is a huge productivity win. The column-chunk storage (.rrd files) is clever—it makes queries and transforms fast without duplicating data. And the direct PyTorch dataloader integration means you can go from recording to training loop without an export step, saving time and compute. The open-source SDK is genuinely free (Apache-2.0/MIT), so small labs and individual researchers can adopt it risk-free. That said, Rerun has a learning curve: its data model and blueprint system require initial investment to get right. For teams that just need a quick chart or a basic dashboard, it's overkill—Foxglove or a simple plotting library is faster. Also, Hub is contact-sales only, which creates uncertainty for teams that outgrow the local SDK and want predictable cloud pricing. Compared to Weights & Biases, Rerun is far more focused on raw sensor data and 3D visualizations, but lacks the experiment tracking and hyperparameter management that many ML teams rely on. In practice, we'd reach for Rerun when you need to deeply understand multi-modal data: debugging a sensor fusion pipeline, inspecting policy rollouts, or visualizing training of a 3D reconstruction model. Where it bites: if your team has no robotics/Physical AI data, the columnar storage and specialized views won't add value over simpler tools. And since the latest releases (0.33.0, 0.33.1) are mostly bug fixes, the core feature set is stable, so early adopters won't face rapid breaking changes.