2026-10-18

Issue #3 (Sunday, Oct 18, 2026)

馃帶 Listen to today's episode 路 6 min

Get this in your inbox every morning.

Good morning, it's Sunday, October 18, and this is The Daily Punchlist. It's a quieter weekend, so today is about the plumbing that makes AI and robots work on real projects. Procore put a new reasoning engine under its AI agents in open beta, and Pirros now lets Claude, ChatGPT and other assistants search a firm's Revit details and families through MCP. On the robot side, Taisei built a system that turns CAD and BIM drawings into robot-ready maps in a few clicks, and researchers shrank an excavator bucket-tracking model enough to run on a small onboard computer. Plus, Bedrock Robotics is reportedly raising $600M, and Brian Potter's essay on where robot training data comes from is your Sunday long read.

In today's issue

1. Procore puts a new reasoning engine under its AI agents

Procore released Magpie 3.0, a new AI reasoning engine behind Procore AI agents, in open beta through Procore Explore. Like Magpie 2.5, it works through multi-step analysis, but Procore says it takes a more direct approach designed to make efficient use of advanced large language models. It only calls Procore and connector actions when a question needs them, which is meant to keep answers focused when a company has many connections set up. Company admins can enroll the whole company or individual projects. Once enrolled, agents set to Magpie 2.5 (or with no model selected) switch to 3.0 automatically. Their "Planning Strategy" setting stops applying, so anything an agent must follow has to move into Custom Instructions. Because answers may differ from what teams are used to, the release note suggests starting with one project.

Why it matters: As more tools plug into Procore through connectors, the engine that decides which one to call matters more. If you try the beta, move your Planning Strategy rules into Custom Instructions before you enroll, or your agents will quietly stop following them.

Source: https://datagrid.com/release-notes/procore-ai-magpie-3-0-open-beta-available-in-procore-explore (Procore AI release notes, Oct 16)

2. Pirros lets Claude and ChatGPT search a firm's Revit details and families

Pirros launched an MCP (Model Context Protocol) connection that lets AI assistants search the details and Revit families a firm stores in Pirros. It works in Claude (including Claude Code), ChatGPT, Cursor and Codex. Users can ask in plain English for typical details and get the PDFs back, or check whether a family already exists before someone builds a new one. It searches by name, category and Revit version and flags near-duplicates. It can also show which families nest a given family, along with its revision history, and it can save picks into collections Pirros calls "stashes." It cannot edit details, families or Revit models. Admins create a credential for each AI tool and choose which workspaces it reaches and whether it can read, write or download. Pirros says customer data never trains the model and that it is SOC 2 Type II compliant. Two caveats from Pirros itself: everyone on a credential gets that credential's full reach, whatever their own Pirros permissions, and detail PDF links an assistant shares are public and don't expire.

Why it matters: Project managers, principals and spec writers who never open Revit can reach firm standards from the assistant they already use. Have IT read those two caveats before you switch it on.

Source: https://engtechnica.com/pirros-brings-firmwide-design-knowledge-into-ai-assistants/ (Oct 16) 路 Pirros announcement: https://www.pirros.com/blog/pirros-mcp (Oct 13)

3. Taisei builds a system that turns CAD and BIM drawings into robot maps

Japanese general contractor Taisei and Kobe-based shop-drawing firm Sun Engineering developed AutoRoboMapSystem (A.R.M.S), a patent-pending system that generates robot-specific maps from AutoCAD (2D) and Revit (3D) data in a few clicks. Some site robots need a simplified digital map to know their work areas and routes. Until now, operators cleaned up drawings by hand, deleting text, equipment symbols and dimension lines one by one, because leftover clutter can make robots malfunction. A.R.M.S uses rule-based automatic processing to avoid missed or wrong deletions, and it reads BIM attributes such as room information and materials to improve how robots recognize their surroundings. Taisei plans to build it into how it runs its construction-support robots.

Why it matters: Robots often stall at setup, not at the task itself. Turning the drawings a project already has into robot-ready maps is the unglamorous step that makes site robots practical.

Source: https://www.taisei.co.jp/about_us/wn/2026/261015_11170.html (Japanese, Oct 15)

4. Altira opens an AI takeoff beta, starting with doors and specialties

Altira, an AI takeoff startup in Plantsville, Connecticut, launched a beta that starts with two CSI divisions: Division 08 (Openings, such as doors, frames and hardware) and Division 10 (Specialties, such as toilet accessories and grab bars). The company says its "Drawing Intelligence" approach reads drawing context and relationships across sheets instead of treating each sheet on its own. The broader platform is meant to handle AI quantity extraction, visual checks against the drawings, revision detection and comparison, quantity change reports and automatic requantification. Altira has also published a "4蟽 Accuracy Framework" for judging how confident estimators can be in a quantity. Beta access is free to register for.

Why it matters: Doors, hardware and specialties are tedious, count-heavy scopes, so they're a sensible narrow place to start. The revision-tracking features are where takeoff tools prove their worth when addenda land.

Source: https://www.prlog.org/13177216-altira-launches-ai-construction-takeoff-beta-starting-with-csi-divisions-08-and-10.html (Oct 16)

5. Researchers track an excavator's bucket even when it's out of sight

A paper in Results in Engineering, by Zhi Chen and colleagues, describes a way to track an excavator bucket's angle when it is buried in soil or hidden behind a truck or material pile. The method combines motion sensors (IMUs) with a depth camera. The team pruned its vision model from 3.08 million to 0.25 million parameters (6.17 MB down to 0.75 MB) while detection accuracy held at 99.4%, versus 99.5% before pruning. It ran on an NVIDIA Jetson TX2 mounted on a compact hydraulic prototype. The typical error (RMSE) was 0.34 degrees with the machine standing still, about 3.9 degrees while leveling, and about 6.4 to 6.8 degrees while digging. The authors say that works out to roughly 1 to 2 cm of error at the bucket. Once trained, the model doesn't need a sensor on the bucket. It is still a lab result: the dataset had 161 samples, augmented to 322.

Why it matters: Bucket-mounted sensors take a beating, so tracking that doesn't need one and runs on small onboard hardware is useful for autonomy retrofit kits. It now needs testing on real machines and sites.

Source: https://scienmag.com/ai-helps-excavators-track-their-buckets-even-when-hidden-from-view/ (Oct 17) 路 Paper: https://doi.org/10.1016/j.rineng.2026.113464

6. Sunday long read: where will robot training data come from?

In Construction Physics, Brian Potter explains how robots get the training data their AI models need. Teleoperation means a person drives a real robot, which gives good data but is expensive. Simulation is cheap but suffers from a "sim-to-real gap" that makes it much harder to use for manipulation than for walking. Recording people with cameras or handheld grippers is cheaper but doesn't map perfectly onto a robot. Some reference points from the essay: Physical Intelligence's Pi 0 trained on 10,000 hours of robot data, and Generalist AI's GEN 1 on about 500,000 hours. Dyna Robotics and Nature Will each claim more than 1 million hours of human video. By one rough measure, even older language models trained on something like 1,000 times more data than today's robot models. Potter's naive estimate for closing that gap is $10 billion to $100 billion. He also notes that data from deployed robots tends to be very repetitive.

Why it matters: Every "foundation model for construction" pitch, like Origin's Construction Action Model from Friday's issue, is really a bet on collecting data. Jobsites full of varied, messy work could become the most valuable training data of all.

Source: https://www.construction-physics.com/p/how-to-train-your-robot (Oct 15)

Trending on X

  1. @EchoCast_, a video that appears to show humanoid robots mixing mortar, buttering bricks with a trowel and laying them between concrete columns: 133 likes, 39 reposts, 38 replies, 8 quotes (9,466 views, 46 bookmarks). Heads up: this video is highly suspected to be AI-generated, and its content is not verified. The post claims a 19-year-old MIT student built the robots for $85,000, but it names no company and we couldn't confirm any of it. It still drew the most engagement of anything on today's list. https://x.com/EchoCast_/status/2111361268284002463
  2. @lukas_m_ziegler, a breakdown of Origin's $20M launch and its OG-1 drywall-finishing robot: 59 likes, 12 reposts, 7 replies, 1 quote (4,523 views). Two days after the announcement, Origin's story is still traveling, now through robotics-newsletter circles. https://x.com/lukas_m_ziegler/status/2111456978333581423
  3. @sshttai, a clip that appears to show a robot in a hard hat carrying lumber across formwork and fitting boards at the slab edge: 18 likes, 6 reposts, 3 replies, 3 quotes (1,120 views). Heads up: this video is highly suspected to be AI-generated, and its content is not verified. No company, site or date is given. The hook ("nobody on the crew said a word") is what got it shared. https://x.com/sshttai/status/2111079953802445227

Tool of the day: Dusty Robotics

If Taisei's robot maps (#3) got you thinking about BIM driving robots, Dusty Robotics is the layout robot in our directory. It prints digital models directly onto the jobsite floor. The company says one person can lay out 10,000 to 15,000 square feet a day with up to 1/16" accuracy. Those are company figures. https://dailypunchlist.com/tools/robotics/dusty-robotics/ 路 https://www.dustyrobotics.com

Funding roundup

Get this in your inbox every morning.