Weights & Biases

Training robotics AI is complex by default because every modality adds another layer of uncertainty. These systems operate in the physical world, generating rich, multimodal outputs like videos, trajectories, segmentation maps, and simulation results across countless experiments. Comparing performance has traditionally been slow, manual, and error prone. As workflows scale, it becomes easier to miss subtle regressions or overlook key insights, slowing progress when speed matters most. To move faster, robotics teams need better ways to evaluate and compare experiments, with advanced visualization and a unified platform to track and analyze model behavior without switching between tools. Weights & Biases brings AI and high fidelity simulation into a single development view, enabling faster iteration, deeper insights, and more rigorous evaluation for physical AI teams. https://loom.ly/f5c_0ro #icra #robotics #WeightsAndBiases

China has a new #1 AI model and its ERNIE 5.1. #13 on the leaderboard. 6% of comparable pre-training. 36 Billion active MOE. Only downside: It's a closed source model. Watch the full episode on our YT to see us dive in deeper. #ernie #china #chinaai #ai #llm

Physical AI pipelines aren’t shaped like the ML training you’re used to. Standard fine-tuning is one shape. Data goes in, model comes out and an eval at the end. A production VLA pipeline is three shapes at once. Data loading runs on CPUs. Training runs on one class of GPU. Stochastic sim-eval runs on another — and a single rollout tells you almost nothing. You need hundreds, with domain randomization, before you can trust that the policy will transfer to real hardware. Each stage wants different infrastructure. Each one scales independently. While all of it has to roll up to a single training decision. On May 12 at 10am PT, our own Anushrav V. from Weights & Biases and Ian D. Jordan, PhD from Anyscale are walking through the actual blueprint: → Orchestrating heterogeneous training + sim-eval workloads with Ray → Aggregating thousands of stochastic rollouts into something you can act on → Tracking video rollouts, trajectories, and task success alongside training metrics in W&B Models → Closing the gap between simulation and real-world deployment If you’re building anything in robotics or physical AI, this one’s worth the hour. Save your spot → https://lnkd.in/gHAHDTgn

Physical AI pipelines aren’t shaped like the ML training you’re used to. Standard fine-tuning is one shape. Data goes in, model comes out and an eval at the end. A production VLA pipeline is three shapes at once. Data loading runs on CPUs. Training runs on one class of GPU. Stochastic sim-eval runs on another — and a single rollout tells you almost nothing. You need hundreds, with domain randomization, before you can trust that the policy will transfer to real hardware. Each stage wants different infrastructure. Each one scales independently. While all of it has to roll up to a single training decision. On May 12 at 10am PT, our own Anushrav V. from Weights & Biases and Ian D. Jordan, PhD from Anyscale are walking through the actual blueprint: → Orchestrating heterogeneous training + sim-eval workloads with Ray → Aggregating thousands of stochastic rollouts into something you can act on → Tracking video rollouts, trajectories, and task success alongside training metrics in W&B Models → Closing the gap between simulation and real-world deployment If you’re building anything in robotics or physical AI, this one’s worth the hour. Save your spot → https://lnkd.in/gHAHDTgn

MCP vs. A2A? The protocol you pick today shapes everything that follows. Join us alongside Weights & Biases, Arcade.dev, FastMCP, and Angie Jones of the Agentic AI Foundation on May 5 for a deep dive on the protocols powering the next wave of AI agents. Register here: https://lnkd.in/dN3UN6vY

Our colleague Russell Ratshin broke a lamp while moving a table to vacuum. 💥 Around the same time, the family decided their living room needed an upgrade — clearing out the kid clutter and turning it into something more grown-up, record player included. You can measure. You can squint and try to picture it. But you don't really know how a sofa, a lamp, or a set of chairs fits your living room until it's there. So rather than keep guessing, Russ built a small AI agent to close that gap — drop in a photo of the room and a photo of the piece, get back a staged image of both together. Then he wrote it up as the inaugural post of our new AI Builders series. The whole thing comes together with a surprisingly small stack: → marimo — the AI-native notebook environment where he prototyped → Google Gemini 3.1 Flash Image Preview, aka Nano Banana 2 — handles the image generation → W&B Weave — wraps the agent with observability, capturing traces, generated images, cost, and latency in one place The post walks through the use case, the setup, the agent code (~50 lines wrapped in a weave.Model class), and why observability and evaluation matter the moment you put an LLM-powered feature in front of real users. 🐈 There's also a brief appearance from Max, Russ's orange cat. If you're building agents — for interior design or anything else — the patterns translate directly. Read the full post: https://lnkd.in/dd6AJgX5 Clone the GitHub Repo: https://lnkd.in/d_eysjr9

Every ML team has the same 3 or 4 links they open on every run. The Grafana dashboard. The Datadog view. The eval results page. The dataset card. Useful, but in W&B they lived inside the config blob. Scroll, find, click. You can now pin selected config keys as References on the run Overview tab. What that means: ➙ Pin one key or many at once ➙ References get their own section at the top of the page, above Notes and Tags ➙ Values render as markdown, so a string like "[Grafana](https://...) | [Datadog](https://...)" displays as a clean row of clickable links ➙ Order is preserved, so you control the layout Small change, real quality of life upgrade if you live in the Overview tab during debugging or reviews. Super easy to get started, see the docs below! #MLOps #WeightsAndBiases #MachineLearning