Runway bets robot policies on video

Today: a video model vendor says robot policies should be pretrained on video rather than demonstrations, two fleet software layers ship, and two open hardware projects put a full autonomy stack within reach of a small team.

Runway ships a policy model that learns from video

Runway announced Praxis-1, its first open-weight world action model, built on the same large-scale video pretraining behind its general world models such as Solaris and GWM Worlds 2. The argument is that a generalist robot policy should inherit physical understanding from third-person video, which is effectively unlimited, instead of from teleoperated demonstrations, which are scarce and expensive. Chief technology officer Kamil Sindi told The Robot Report that most robot policies are bottlenecked by robot data, and that scaling video raises the ceiling instead.

Two numbers stand out. Runway says simulating robot policies inside its world model predicts real-world results with 0.95 correlation, and in its own placement-error comparison a policy pretrained on web video reached 16.1 centimeters final placement error after fine-tuning, against 16.0 centimeters for one pretrained on teleoperated robot video, with error bars of plus or minus one standard error over 93 evaluation pairs, so the two data sources are not distinguishable in that test. Runway’s demos target what defeats demonstration-trained policies: clutter, transparency and deformable items with no fixed grasp point.

Early partners are Noble Machines with bimanual manipulation, Standard Bots with its 6 degree of freedom RO1 arm, and Ultra on a mobile base, with an open-weight public release planned for the coming months. The interesting claim is not that video beats teleoperation yet, it is that the fine-tuning curve is steep enough that a few hours of your own data may be enough, which is the argument behind why vision-language-action models matter and the robot data bottleneck.

NVIDIA adds a 64 GB DGX Spark for local inference

NVIDIA expanded its DGX Spark line with a 64 GB unified memory configuration, per its developer blog. It keeps the GB10 Grace Blackwell superchip, DGX OS and the CUDA software stack of the larger model, and NVIDIA says it runs models up to 100 billion parameters on device, or 200 billion across two units linked over ConnectX-7 that pool their memory to 128 GB. NVIDIA claims up to 1.7 times the performance of a single node in its Qwen 3.8 27B test.

The software list matters more than the silicon: Agent Toolkit, CUDA-X libraries, Nemotron open models, Ollama, vLLM, PyTorch and llama.cpp, with a Sync Cluster Assistant that configures connected units. For a robot team this is a workstation, not a robot computer, since inference inside the machine still belongs on something like a Jetson, the choice covered in our Orin Nano versus Orin NX guide. It shortens the loop around the robot: policy evaluation, dataset processing and fine-tuning on hardware next to the machine instead of an hourly cloud instance.

ANYbotics turns ANYmal data into work orders

ANYbotics launched Shift, a cloud platform in four modules. Shift Fleet schedules and monitors a robot fleet and can be triggered by a distributed control system. Shift Insight combines thermal, acoustic, visual and gas readings with operating conditions such as motor speed, load and ambient temperature, and only calls a reading anomalous when the context says it should. Shift Connect turns confirmed anomalies into work orders through connectors for Yokogawa, SAP, IBM Maximo, GE Vernova APM, Cognite, Oracle, Siemens Energy and SLB, plus a REST API documented with OpenAPI, while Shift Maps ties a plant model to asset IDs.

Shift ingests inspection data from third-party robots today; full fleet management for non-ANYmal hardware starts next year, and the company says no unified inspection robot standard exists in the US or Europe, since VDA 5050 was written for intralogistics AMRs. Its downtime figures, 1.4 trillion dollars a year for the 500 largest companies and 250,000 dollars an hour per outage, are its own estimates. For integrators the value is a documented API where the guard layer in a modular robot stack usually has none.

ASI and SoftBank take autonomy to heavy equipment

Autonomous Solutions Inc. and SoftBank announced a joint venture to commercialize brand-agnostic autonomous heavy equipment, with SoftBank providing 225 million dollars in new funding to ASI, per The Robot Report. ASI has built the Mobius fleet orchestration system since 2000, and chief executive Mel Torrie frames the problem as mixed fleets on one jobsite: haul trucks, dozers, loaders and compactors from different manufacturers under one coordination layer.

That OEM-agnostic framing mirrors the interoperability argument in logistics AMRs, and it is harder here because construction machines are hydraulically actuated and sites are unmapped. It is the industrial end of the curve described in five million robots and the missing refusal layer.

An open 12-DOF quadruped that runs the full Nav2 stack

A solo builder published Orion, a 12 degree of freedom quadruped designed, built and programmed over eight months, now at 268 stars on GitHub. The stack is unusually complete for a personal project: an NVIDIA Jetson Orin Nano runs an Isaac ROS Docker workspace with TensorRT, dual CSI cameras feed stereo depth through MiDaS and SGBM, an RPLIDAR A1M8 drives slam_toolbox and Nav2 for mapping and point-and-click pure pursuit navigation, and FreeRTOS firmware on an STM32F401 handles inverse kinematics, walking gaits and IMU stabilization on custom KiCad control and power boards. Training uses PPO in Isaac Lab across flat and rough terrain from a SolidWorks to URDF to USD pipeline.

If you are assembling a first ROS 2 platform, this repository is a checklist of the layers usually skipped, and it pairs with what Isaac Sim does on the simulation side. The repository does not declare a licence, so treat it as a reference rather than a base for redistribution.

OpenArm 2.0 adds an in-hand camera and a leader arm

Enactic, based in Tokyo, released OpenArm 2.0, an open hardware 7 degree of freedom arm for physical AI research, teleoperation and contact-rich data collection. Each arm carries eight DAMIAO actuators in a quasi-direct-drive, backdrivable arrangement (two DM-J8009P at the shoulder, one DM-J4340P and one DM-J4340 mid arm, four DM-J4310 for the wrist and gripper), a 1 kilohertz control loop on CAN-FD with classic CAN fallback, and dual 14 bit magnetic encoders per motor. Payload is listed at 4.1 kilograms nominal and 6.0 kilograms peak, with 606 millimeters of reach.

New relative to the version we covered in September are the in-hand RGB camera on the two finger gripper and two ecosystem additions: OpenArm KER, a motorless leader arm with magnetic encoders for low resistance teleoperation, and OpenArm Cell, an evaluation rig of roughly 100 kilograms with a 300 millimeter Z-axis lift and light curtains. Software covers ROS 2, MoveIt 2 bimanual configurations, Dora nodes, MuJoCo, Isaac Lab, LeRobot and mink, with CAD under CERN-OHL-S-2.0 and most code under Apache 2.0. Enactic works through certified manufacturers rather than selling directly, and the KER timeline is still to be announced.

Quick Bites

• Clearpath open sourced its ROSCon 2026 workshop on ROS 2 middleware, four labs covering Fast DDS, Cyclone DDS and Zenoh tuning, packet capture with Webshark, stress testing under a shared network bottleneck and MCAP replay benchmarking (GitHub). See also our QoS guide.
• Innodata opened a lab that captures 3D motion directly from human and mechanical bodies for humanoid policy training, and will independently validate performance figures robots report about themselves (The Robot Report).
• The Open Robotics annual survey for 2026 is open, and the project notes it collects no telemetry, so this is the main feedback channel for ROS, Gazebo, Open-RMF and ros2_control (Open Robotics Discourse).
• A ROS 2 Lyrical sync is planned for 5 October, with 307 packages waiting and 3 regressions reported (Open Robotics Discourse).
• NVIDIA says GPT-6 Astra Ultrafast generates tokens up to 8 times faster than Astra Standard on Blackwell GPUs, which shortens the edit, test and debug loops that coding agents run on robot software (NVIDIA).