Grabette, LingBot-VLA 2.0 and LeRobot 0.6 | AI robotics

What do I take from these two weeks if what I want is to build robots on my desk? Not humanoids. What actually changed is that you can now generate your own manipulation data without owning an expensive arm, and train with open weights on a gaming GPU. Straight to the point.

Grabette: record data with your hand, no robot needed

Pollen Robotics published Grabette on 21 July, a gripper you hold in your hand that records what you do: you go around picking things up and out comes a dataset in LeRobot format, ready to train on.

The numbers they give: around 490 € of bill of materials for the Grabette and around 120 € for the Gripette, the version that mounts on the robot. All Apache-2.0, and not just the software: the CAD, the production files, the capture and the post-processing. Inside it runs a Raspberry Pi, a Pi camera, an OAK-D depth camera, magnetic encoders and an IMU. Zero proprietary parts.

Why does this matter more than any humanoid headline? Because it closes the loop at home. Before, to get manipulation data you needed the robot and you had to teleoperate it. Now you record with your hand, you train, and then you move it onto the hardware. In other words: data stops being the bottleneck.

LingBot-VLA 2.0: 6 billion parameters that fit on your 4090

Robbyant (Ant Group) released LingBot-VLA 2.0, a 6B-parameter vision-language-action model with a Qwen3-VL-4B-Instruct backbone, trained on roughly 60,000 hours of data (50,000 of robot trajectories across 20 different setups and 10,000 of egocentric human video). Unified 55-dimensional action space: arms, gripper, hands, waist, head and mobile base.

The number you care about: the authors report roughly 130 ms per inference call with 10 denoising steps on an RTX 4090D. Careful, that figure comes from the paper itself, not from an independent measurement, so treat it as a reference and not as a promise. The licence is clear though: Apache-2.0, weights included.

Translated: a generalist VLA you can fine-tune for your own arm without asking anyone for permission or renting a cluster.

LeRobot 0.6.0: update carefully

If you have a project sitting on LeRobot, read this before you run pip install -U. Version 0.6.0 (6 July) brings new policies, six simulation benchmarks under lerobot-eval, a lerobot-rollout CLI, FSDP, a dataloader twice as fast, native support for the Seeed Studio reBot B601-DM arm and better SO-100/SO-101 support. On the same day, NVIDIA announced Isaac GR00T 1.7 and Isaac Teleop landing in LeRobot.

And now the part that breaks things:

• Minimal install: training dependencies moved into the [training] extra.
• PyTorch is now 2.7 or higher.
• sac is renamed to gaussian_actor.
• eval_freq becomes env_eval_freq.
• --dataset.vcodec becomes --dataset.rgb_encoder.vcodec.

If your script has been running for months, any one of those five will leave it dead. Don’t overthink it: pin the version, update on a separate branch and test.

A real quadruped running with onboard perception

Science Robotics published APT-RL, a framework that learns locomotion representations from trajectory optimisation and uses them as a starting point for reinforcement learning. KAIST’s HOUND quadruped cleared stairs, fences, stepping stones, gaps and fallen branches, with instantaneous peaks of up to 6 m/s during a drop manoeuvre, using onboard compute and perception only (depth, 2D LiDAR and proprioception).

What matters here for you isn’t the robot, it’s the method: they generated 180,000 optimisation trajectories in 8 minutes and used them as a prior, instead of training RL from scratch. That pattern scales down to small and cheap.

Vision on a Raspberry Pi, no accelerator

On 15 July Raspberry Pi published an official guide, written with people from Intel and Ultralytics, for running YOLO26 models (the nano variant) on a Pi 4 or Pi 5 with the OpenVINO runtime, on CPU only. No Hailo, no Coral, nothing to buy.

An honest warning: the post shows a comparison chart but gives no FPS or latency figures in the text. So don’t trust any number you see out there citing that article. If you need one, build it and measure it yourself.

What did NOT happen

Two weeks without a single robotics foundation model release from Physical Intelligence or Google DeepMind. The latest from PI is April (π0.7) and the latest from DeepMind on robotics, Gemini Robotics-ER 1.6, is also April.

And the flashiest headline of the fortnight comes from a closed system: Sunday Robotics says it folded 778 out of 785 garments correctly and autonomously, 99.1 %, in homes the model had never seen. That figure is published by the company itself, with no independent verification, and neither the model nor the data nor the hardware are open. I’ll leave it there, as a contrast: the pretty number lives in the closed system, and what you can replicate at home is in the first three items of this entry.

What I’d do this week

If you’re starting out, forget the VLA. Build the obstacle-avoiding robot, understand why the wiring works, and then move up. If you already have an arm, look at Grabette: recording your own data is what will set you apart, not the model you pick. And if you have a project on LeRobot, go through those five breaking changes before they catch you mid-training.

I know that all at once it feels like a lot. Relax: pick one of these five and do that one. The rest will still be here next week.