Weekly Robotics #357

We are again happy to be the media sponsor of Cracow Robotics and AI meetup. Can't wait!

Today's issue starts off very VLA/general-robots-heavy, since there have been quite a few interesting developments in recent days. Then we will go into some cool projects and end with some robot fails. Enjoy!

This issue is brought to you by our sponsors:

Multiply Labs Advances AI for Biologics Robotics on Anyscale [Sponsored]

Multiply Labs Advances AI for Biologics Robotics on Anyscale cover

Multiply Labs is using advanced robotics to remove human-based contamination from biologics manufacturing. To drive coverage for more complex edge cases, the team needed to scale their training runs across multiple clouds - AWS, GCP and Nebius - to gain flexible, on-demand access to large GPUs (H100s and A100s). Learn how they achieved this with Ray - the world's most widely used engine for AI - on Anyscale.


mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs

mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs cover

Please allow me to quote the abstract instead of summarizing:

we introduce mimic-video, a novel Video-Action Model (VAM) that pairs a pretrained Internet-scale video model with a flow matching-based action decoder conditioned on its latent representations. The decoder serves as an Inverse Dynamics Model (IDM), generating low-level robot actions from the latent representation of video-space action plans. Our extensive evaluation shows that our approach achieves state-of-the-art performance on simulated and real-world robotic manipulation tasks, improving sample efficiency by 10x and convergence speed by 2x compared to traditional VLA architectures.

To learn more about this work, check out this paper.


Gemini Robotics ER 1.6: Enhanced Embodied Reasoning

Gemini Robotics ER 1.6: Enhanced Embodied Reasoning cover

The team at DeepMind released Gemini Robotics ER 1.6, which is available to developers through the API. I really liked the gauge-reading examples, and I hope to see some developers apply this in their projects!


A Steerable Model with Emergent Capabilities

A Steerable Model with Emergent Capabilities cover

Physical Intelligence has released a new model called Pi0.7. This time, the team created a model that can generalize and compose behaviors - being able to execute tasks that the model was not explicitly trained on. I have one question, though - is this how you are supposed to cook sweet potatoes in an air fryer?


Chinese robot smashes human world record in half-marathon: ‘Just whooshed right past me’

Chinese robot smashes human world record in half-marathon: 'Just whooshed right past me' cover

The droids went racing in China, this time beating a human record for a half-marathon, completing the 13-mile race in 50 minutes and 26 seconds.


UnrealRoboticsLab

UnrealRoboticsLab cover

URLab (Unreal Robotics Lab) is an Unreal Engine 5 plugin that embeds the MuJoCo physics engine directly into the editor and runtime. Drag-and-drop MJCF XML import, a component-based architecture that maps 1:1 to MuJoCo elements, the full MuJoCo C API accessible from C++ and Blueprints, ZMQ networking for external control, Python policy integration, 40+ sensor types, 8 actuator types, debug visualization, and a record/replay system.


The Hidden “Hand Farms” of India: Fueling the AI Robot Revolution with Human Motion

<img src="https://www.weeklyrobotics.com/img/features/357/The_Hidden__Hand_Farms__of_India__Fueling_the_AI_Robot_Revolution_with_Human_Motion.jpg" alt="The Hidden "Hand Farms" of India: Fueling the AI Robot Revolution with Human Motion cover">

When I used to think of general robots, the comparison with autonomous cars always came to mind, but what I failed to factor in was that: a) it’s cheaper to collect this data, and b) it might be much easier to scale. This article will give you some insights into how datasets like these can be captured, including the ethics of such operations.


Robot golf vs holes that keep getting harder

Robot golf vs holes that keep getting harder cover

I find it incredible how good a result Stuff Made Here was able to achieve with a one-degree-of-freedom robot shown in this video. I highly recommend checking it out if you like seeing engineering applied to not-so-serious projects.


All the Mini Cheetah Fails

Ben Katz had published a compilation of the Mini Cheetah quadruped failing. Seeing this reminded me that robotics is actually properly difficult.


Delivery robots having a hard time

Speaking of fails, this Reddit compilation proves my point. I wonder how often the robots with a flag have close calls at intersections. Does anyone have any statistics they can share?


Our Sponsors

  • Anyscale gives you a platform to run and scale all your ML and AI workloads, from data processing to training and inference.
  • Foxglove is a purpose-built, modular platform for robotics teams to collect, organize, and learn from vast quantities of multimodal data, creating the data flywheel to safely scale from development to distributed fleets
  • HelloRobo distils robotics tech into simple, usable interfaces
  • Jiga connects hardware teams directly with vetted manufacturers for reliable capacity from prototype to production, combining the speed of digital manufacturing with the trust of a long-term supplier relationship

Events

For more robotic events, check out our event page.


Want to promote your product or service in Weekly Robotics? Check out our advertising options.

issue 356