Perseverance record and Autoware research center

Yesterday’s roundup covered edge vision models and Digit 5 charging. Today: a rover that sets its own record, an open source autonomy lab in Munich, a safety simulation startup, and natural language programming for industrial arms.

Perseverance sets the off-world driving record

NASA’s Perseverance rover became the farthest driven vehicle on another world on 11 October, when a 299.6 foot (91.31 meter) drive pushed its odometer to 28.10 miles (45.22 kilometers) (JPL). That passes the 28.06 miles (45.16 kilometers) logged by Opportunity between 2004 and 2018, and it took about five and a half years rather than almost 15.

The number is really an autonomy result. JPL says more than 90 percent of the rover’s total distance was driven using AutoNav, against roughly 5 percent for the first AutoNav on Spirit and Opportunity. The system images terrain while rolling, builds a hazard map, and plots a course between waypoints planners on Earth select, a mode the team calls thinking while driving, with a best single sol of 1,351 feet (412 meters). The perception budget grew with it: Sojourner checked 20 measurement points per look, Perseverance processes 600,000 (IEEE Spectrum).

The rover receives goals rather than paths, the same split between planner and local controller that ROS 2 navigation stacks implement. The remaining bottleneck is perception: Mark Maimone, deputy team chief for robotic operations, says glare in the navigation cameras is why the rover parks by mid-afternoon when driving toward the Sun.

TIER IV and TUM put a safety guardian around learned driving

TIER IV and the Technical University of Munich established the TUM Autoware Research Center by TIER IV on 15 October, a multi year collaboration across three TUM groups: the Chair of Automotive Technology, the Professorship of Cyber Physical Systems and the Professorship of Autonomous Vehicle Systems (TIER IV).

The agenda is the interesting part. Under safety, the center will build a hybrid architecture that pairs deep learning driving models with rule based supervision, including a verifiable safety guardian that monitors behaviour and intervenes when defined constraints are violated. That is the propose and dispose pattern moved into a vehicle, with the learned model proposing and a separate, checkable component holding veto power. Under evaluation, the plan is a benchmarking framework spanning simulation, offline analysis and real world operation, built on TUM’s EDGAR research vehicle and its CommonRoad scenario tooling.

TUM has already shown that process can be fast: the team deployed EDGAR and demonstrated autonomous operation on Berlin roads within a week, working with TÜV-Verband on an assessment methodology for Level 4 driving for the German government. A funded European centre that produces test methodology and upstream patches matters more than another demo video.

A robotaxi petition tests the school zone edge case

NHTSA began reviewing a defect petition against Tesla’s robotaxi fleet on 16 October, after the request arrived in late September (ExtremeTech). The petition alleges that roughly 1,000 Model Y robotaxis and Cybercabs exceed posted speed limits in active school zones, and reportedly rests on three video links tied to four incidents: one complaint filed with the agency and three social media reports. It does not claim any crash, injury or death.

Keep the status straight. A petition review is not a finding. Anyone can file one, and the agency has 120 days to decide whether to open a formal investigation, which is the only step that can lead to a recall. NHTSA previously examined Full Self-Driving over traffic violations and is reviewing certification data for the Cybercab, which has neither steering wheel nor pedals. Waymo has faced inquiries over stopped school buses and a collision with a child outside a California elementary school.

For AV developers the lesson is about evidence. School zones concentrate the hardest parts of an operational design domain: children, occlusion, speed limits that change by time window, and inconsistent roadside behaviour. Fleets need speed and zone compliance logged in a form an investigator can replay, the same observability argument that applies to any robot running near people.

SafeWorld raises 12.2 million dollars to test robots around people

SafeWorld came out of stealth this week with 12.2 million dollars in seed funding, co-led by Shine Capital and a16z Speedrun, with the Carnegie Mellon University Endowment among the backers (The Robot Report). The Oakland company sells simulation software for safety cases: teams build scenarios in a browser from past incidents, safety standards and robot logs, then run the robot through thousands of variations with reactive human motion and measure safety performance.

Co-founder Ding Zhao, who directs the Safe AI Lab at Carnegie Mellon, frames the shift as robots leaving cages and sharing space with people, and CEO Kyle Wong argues scenario libraries should not be rebuilt per customer. SafeWorld says customers span industrial, manufacturing, logistics, construction and some home use, and that tests can be rerun as software changes. The architectural point stands regardless: if safety evidence is regenerated on every software change, it belongs in the same pipeline as the build, close to the simulation stack and behind the same guard layer that limits the policy at runtime.

Standard Bots teaches industrial arms by description

Standard Bots released three AI capabilities for its robot platform: Agent, Locate and Inspect (ASSEMBLY). Agent takes a natural language description, asks clarifying questions, and builds the program with the operator. Locate uses the robot’s camera feed so a user can tap a part or describe it instead of teaching a grid of waypoints, which the company says cuts programming time by 90 percent. Inspect, built with Roboflow, runs vision models trained on a manufacturer’s parts, and customers can load their own.

Leif Jentoft, head of applied AI at Standard Bots, frames the target as the integrator bill: months of tuning, and a program nobody on site can fix at night. The percentages are vendor claims with no independent measurement, so read them as directional. Teaching by object reference rather than by pose is the simplification that VLA policies promise, delivered as a feature on a conventional arm, and it pays most in high mix lines, where changeover cost rather than cycle time keeps small batches manual.

ROS 2 gets terminal tools and a tool index

The ROS news for the week of 12 October collects several small developer releases worth knowing (ROS Discourse). terminal_rviz renders a stripped down RViz view in a terminal, which solves the SSH session without X11 forwarding (GitHub). DendROS colours ROS 2 terminal output and bundles small utilities (ROS Discourse). robodev.tools indexes a few thousand ROS tools found on GitHub, an attempt to stop generated tools from rebuilding what already exists (robodev.tools).

Also on the list: multi_drone_slung_load_master for slung load carrying with a multi drone system (GitHub), Isaac ROS 5.0 packaged with Pixi, EasyNav 0.5.0 for Humble through Lyrical, and a Zenoh network explorer (GitHub).

Quick Bites

• UNCAP lets connected vehicles swap short natural language descriptions of what they see, weighted by calibrated confidence. In CARLA on the OPV2V benchmark the authors report 63 percent less bandwidth, a 31 percent higher driving score, 61 percent lower decision uncertainty and four times the collision margin in near misses, and sharing raw images did worse than sharing language (arXiv:2510.12992, Robohub).
• Google DeepMind used Gemini 4 Argon to watch a human play cornhole, build a MuJoCo simulation from the video, and compute the control for Apptronik’s Apollo 2 humanoid to sink seven consecutive real world shots (IEEE Spectrum), a follow on to the Argon launch.
• TAKTO is an open source hand exoskeleton that measures finger position with magnetic encoders and adds actuation for force feedback, with build documentation and a digital twin console (Hackaday, GitHub).
• Symbotic acquired Bremen based cellumation, whose celluveyor replaces fixed conveyors with a grid of individually driven wheels for sorting parcels. The brand and headquarters stay, and Symbotic will fold the hardware into its warehouse systems (cellumation).