Bedrock raises $600M for driverless excavators, threatening construction's proud tradition of finishing late

Bedrock Robotics raises $600M for excavators that dig with no one in the cab

The Series C, co-led by CapitalG and Emergence, values Bedrock at $4.1B (@Techmeme) and brings its total funding to nearly $1B; next it moves beyond excavators toward coordinated fleets (@BedrockRobotics). Since August, excavators fitted with its kit have done paid, fully autonomous work on live commercial job sites (@BedrockRobotics). The kit retrofits machines contractors already own, and six customers have deployments underway (@roboindexio). The valuation is up from $1.75B at its Series B in February (@boishivers).

Applause for a robot that already earns money, and little debate. @AnandButani told the founder's story: "In April 2019, @bsofman's robot company burned through $200 million and laid off every employee. This month investors valued his new one at $4.1 billion." CEO Boris Sofman's pitch is the night shift: "fuel up a machine at the end of the day and let it go an extra shift overnight, and now you've boosted your productivity by oftentimes 60, 70, 80%" (@MTSlive).

Sofman's best number is about interest: a developer paying 9% on a $400M loan saves $6M by finishing two months early. A lot of the people borrowing like that today are building data centers, so excavators that dig all night get new compute online sooner (Alphabet, which builds plenty of them, has now led two of Bedrock's rounds).

1. Tesla widens its Robotaxi service area in Houston

"Houston, we have liftoff. Bigger service area now live", with a new geofence shaped like a Starship (@TeslaNewswire). Later the same day: "Austin's robot population is growing" (@robotaxi).

Fans see the pace picking up: "Uh yeah Robotaxi scaling is getting insane" (@DirtyTesLa). Tesla's AI head @aelluswamy: "Also a very useful geofence" (57K).

Heat: high · 3.5M views, 25 posts · quiet on Reddit

2. Latvia approves Tesla's FSD (Supervised), the 10th European country to do so

Latvia's road authority recognized the Dutch RDW's approval, so all three Baltic states now allow it, and the rollout begins soon (@FSDEurope). In Europe it's called Tesla Assisted Driving, and EU-wide approval needs 15 countries covering 65% of the EU's population (@piloly).

Celebration, led by @elonmusk: "Supervised FSD approved in Latvia! 🇱🇻" (1.2M). The catch, per @piloly: "It's another small country, but it still brings us one vote closer to the 15 needed." @EVStraya: "Germany, France and Italy are still waiting".

Heat: high · 3.3M views, 21 posts · quiet on Reddit

3. GXO takes Agility's Digit humanoids to two more US warehouses

Deployments in Statesville, North Carolina and Lebanon, Tennessee start in early 2027, after Digit moved more than 100,000 totes at GXO's Flowery Branch, Georgia site under a 2024 robots-as-a-service deal. The same day, Agility said Digit 5, unveiled in September, has over $300M in committed multi-year orders, and that more than 30 prospective customers are in talks (Agility).

Little so far beyond Agility's own line, that "repeat customers are the signal that matters" (@agilityrobotics).

Heat: low · 59K views, 17 posts · quiet on Reddit

4. CircuitHub's The Grid builds circuit boards in Massachusetts, twice as fast as Shenzhen

▶ Watch the video

A factory-scale robot that builds boards on demand, prototypes or production batches, with no changeovers and no per-design setup. In CircuitHub's own comparison, five four-layer motor-control boards cost $630 and arrive in 4 days, against $465 and 9 days from a Shenzhen shop, shipping and tariffs included (@Michaelzsguo).

Five robotics accounts on our roster quoted it, mostly cheering: "In your face, Shenzhen!" (@Abhindas1), "more factories!" (@aphysicist). @mikekalilmfg: "The component library seems like a bigger deal than the robotics angle, with 35,000+ common parts stocked."

Heat: medium · 310K views, 24 posts · quiet on Reddit

5. John Carmack: no robot can change the batteries in a jack-o-lantern

His reality check on "flashy robot demos": a robot might fetch the pumpkin and open the drawer, but opening the battery pack, finding a precision screwdriver and undoing four tiny screws are still beyond any robot. A specialist robot could be trained for it, he says, but replacing "the full range of human capabilities" is a long way off.

Musk agreed: "These are real benchmarks" (@elonmusk, 237K). Then he raised the bar: "How about the robot is looking down while assembling a Swiss watch with one hand and then, without looking up, catches an oblique 100mph baseball with the other hand?" (@elonmusk, 241K).

Heat: medium · 843K views, 13 posts · quiet on Reddit

AW-0: one model that writes the robot's plan as code, predicts the video and outputs the actions

Trevor Darrell and colleagues built a world action model that generates skill-calling code, future video and motor actions together, so the plan shapes both the prediction and the control. It was pretrained on about 3,000 hours of robot and human demonstrations and adds 40 ms of latency over the base model on an RTX 5090. It matters because it puts reusable, inspectable skills inside a world model instead of bolting an agent on top.

DexPort: a handheld rig with a real robot hand for collecting dexterous demos

Amber Xie and colleagues put a real robot hand in a handheld interface, so demos record the hand's joint states and commands with no embodiment gap. Recording the commands, not just where the fingers stop, raised average success from 11% to 64% across three physical tasks (@dustinturnerai). It matters because dexterous-hand data no longer needs a full teleop setup.

WorldOracle tests whether robot world models predict what really happened

AGIBOT and NVIDIA, with Princeton, UT Austin and the University of Toronto, compare world-model rollouts with real executions across 335 episodes (in the human-operated ones, about 40% of segments are failed attempts), on four measures from outcome to contact detail (@sciencegirl). It matters because a world model that shows a failed grasp succeeding is useless for evaluating policies.