The easiest way for agents to use your robotics data 🤖
Sydney, Australia
Joined August 2025
- Tweets87
- Following25
- Followers1.7K
- Likes97
Pinned Tweet
Robotics teams are drowning in their own data.
Terabytes of images, logs, and 10 kHz time-series.
99% of it never gets looked at.
Incident reports take days.
Root-cause analysis is manual.
Most teams burn 50–80 % of engineering time just trying to find the failure.
We built Alloy to end that.
Alloy is the first platform that lets you search years of robot data in natural language, in seconds.
“Show me every gripper slip when IMU Z-accel > 4 g in salty air" → exact video clips + sensor slices + logs instantly.
No labeling. No scripting. No 1.3 GB MCAPs in Slack.
Tesla & Amazon built this internally.
We're bringing that level of observability to the rest of the world.
We've raised $4.5M pre-seed from Blackbird, Airtree, Skip, Xtal and angels at Tesla, Waymo, OpenAI.
And we're just getting started.
"Mini Cheetah, the Falling Cat" (Kurtz, Li, Wensing, Lin, ICRA 2022) gets a robot with no spine to land on its feet like a cat.
- In the air, total angular momentum stays at zero
- No spine to twist, so the legs do it: they sweep one way and the body turns the other
- The legs were too light, so the team added 3D printed boots with two rolls of nickels each, about 500 g per foot
- They solved 400 optimal falls offline with DDP, then trained a 2x512 network on the answers
- It plans its recovery in 2 ms and lands on its feet.
Robotics problem solved: LeetCode 1041, reported as an Amazon interview question, run in a small Python sim.
Two robots repeat one list of moves forever:
1. go, go, left, go: back home after four runs
2. go, go, left, go, right: walks off forever
How do you tell from the list alone?
we're sponsoring Long Horizon, the physical AI summit by @encord_team 🤖
find us at booth ES4, Fort Mason, Oct 20-21. we'd love to hear how you're scaling your fleet.
kicking it off with us tomorrow night? happy hour with Antioch: luma.com/9urv7cgb
Two kinds of robot training data, according to Amplibotics AI founder James Wigglesworth:
- Egocentric: a GoPro on someone's forehead doing factory work. Not constrained by the robot, but really good for pre-training.
- Teleop: real robots driven by people, recording camera images alongside joint positions and control commands.
His take: teleop is the gold standard. A lot of the panelists at Actuate want more of it, but they can't afford it or there isn't enough. Amplibotics' mission is that volume and scaling problem.
This dexterity paper visualised: "DiSECt: A Differentiable Simulator for Parameter Inference and Control in Robotic Cutting" (Heiden et al., RSS 2021) teaches a robot arm to saw.
- Fit a simulated banana until its knife force matches the real one
- Hold the cut together with springs that weaken smoothly, so gradients exist
- Optimise 5 knife keyframes to minimise mean force, starting from a straight cut
- Give the blade ends, and the sideways wiggle becomes sawing
Real Franka, real banana: mean knife force down 46.5%.
Explaining SwarmFarm's patent, "Coverage path planning" (US 12,093,049 B2, Fitch et al., 2024) that turns an awkward paddock into a sequencing problem.
- Inputs: field boundary, obstacle and no-go polygons, AB line heading, existing wheel tracks, implement width (sets the pass spacing)
- Cellular decomposition: split the non-convex field into cells around the obstacles
- Boustrophedon coverage inside each cell, passes parallel to the AB line
- Optimise the cell sequence and the transitions between cells: every metre from one cell's exit to the next one's entry is non-working travel
Kicking off Long Horizon with @antiochrobotics.
Antioch tests robots in sim before they ship. Alloy finds what went wrong once they're in the field. Come argue which is harder, over drinks.
Monday, Oct 19, 6:30pm, Marina district, SF
Request a spot: luma.com/9urv7cgb
This paper visualised: "Robust Physical-World Attacks on Deep Learning Visual Classification" (Eykholt et al., CVPR 2018) turns a stop sign into Speed Limit 45 with four black and white stickers.
- Optimise the stickers by gradient descent on the image, averaged over real photos at many distances and angles
- Run an L1 pass over the whole sign first to find where it's weakest, then put the stickers there
- 100% targeted success in the lab, 84.8% of frames from a moving car
- The target was LISA-CNN, a research sign classifier, not a self-driving stack
Then we undo it with PatchCleanser (Xiang et al., USENIX Security 2022): mask the sign piece by piece and find the view the stickers can't fool.
This perception paper visualised: "Phantom of the ADAS" (Nassi et al., ACM CCS 2020) made a driver-assist car stop in the middle of the road for a stop sign that was on a billboard for 125 ms.
- Hide a stop sign in three frames of a video ad
- The car's camera believes anything sign-shaped that lasts a few frames
- Projected signs work too
The fix, GhostBusters: four small networks check the sign's light, context, surface and depth, and a fifth makes the call. It caught 99.4% of phantoms with zero false alarms on real signs.
"Hardware Design for Table Tennis Robot Capable of Beating Professional Players" (Mukai et al., Sony AI, arXiv 2606.26643) is Ace's build sheet.
Body: 8 DoF, a 6-DoF arm riding an XY-stage.
Eyes: three event cameras behind swinging galvo mirrors, zoomed in through telephoto lenses.
Event pixels fire only when brightness changes, so the spinning logo draws an arc. That arc gives the spin axis and speed.
The match results are in the Nature paper (Dürr et al., 2026): over 75% returns on spin up to 450 rad/s, and 3 of 5 matches won against elite players.
we're sponsoring the first @TheHumanoidHub conference in SF 🤖
day 2 today. come find us on the demo floor at the AWS Builder Loft.
Alloy Robotics retweeted
had fun chatting with @IlirAliu_ on how we scale robots beyond demos
all the magic (and challenges) are in deployments!
99% robot reliability sounds impressive. Until you scale to 10,000 actions a day.
@_joe_harris_ , Founder & CEO of @AlloyRobotics, explains why reliability becomes one of the biggest challenges as robotics companies move from prototypes to real-world deployments.
His company is building the AI Robotics Engineer, using AI agents to investigate failures, analyze robot data, and help engineering teams operate larger fleets.
Founded in Australia, Alloy has raised A$16M+ from Square Peg, Blackbird, Airtree, and Skip Capital, and now works with companies including Andromeda, Advanced Navigation, and SwarmFarm Robotics.
Joe previously helped scale Eucalyptus, which was later acquired in a deal valued at over $1 billion.
In our latest @22Astronauts_ podcast, we discuss robotics infrastructure, raising venture capital, China's role in robotics, the future of software, and what it takes to build a company in Physical AI.
This episode is brought to you by @CoreWeave, providing the specialized cloud infrastructure powering physical AI and large-scale model training.
𝗧𝗜𝗠𝗘𝗦𝗧𝗔𝗠𝗣𝗦
00:00 Building the AI Robotics Engineer
03:25 Why customers buy solutions, not platforms
09:30 Scaling Eucalyptus and lessons from startups
17:30 Founding Alloy Robotics
22:45 Raising venture capital and startup storytelling
28:00 Building products customers actually want
36:30 China's robotics ecosystem and global supply chains
44:00 AI coding and the future of software
47:00 Build vs. buy in robotics
51:11 Why 99% robot reliability isn't enough
Full Episode:
🎥 YouTube: youtu.be/ufa9lJ01mF8
🎧 Spotify: open.spotify.com/episode/0k3…
Recreate lost gyro data with MoE-Gyro. A self-supervised deep-learning framework designed to solve the fundamental trade-offs in MEMS (Micro-Electro-Mechanical Systems) gyroscopes between measurement range and noise reduction.
MoE-Gyro (Pan et al., NeurIPS 2025) learns to rebuild the clipped part of a gyro reading. We built a gentler swing at the same 2.5 Hz in our own sim and trained the authors' model on real iPhone gyro recordings. The heading added up from the clipped reading ended 22° off. From the rebuilt reading, it ended under 2° off.
Alloy Robotics retweeted
99% robot reliability sounds impressive. Until you scale to 10,000 actions a day.
@_joe_harris_ , Founder & CEO of @AlloyRobotics, explains why reliability becomes one of the biggest challenges as robotics companies move from prototypes to real-world deployments.
His company is building the AI Robotics Engineer, using AI agents to investigate failures, analyze robot data, and help engineering teams operate larger fleets.
Founded in Australia, Alloy has raised A$16M+ from Square Peg, Blackbird, Airtree, and Skip Capital, and now works with companies including Andromeda, Advanced Navigation, and SwarmFarm Robotics.
Joe previously helped scale Eucalyptus, which was later acquired in a deal valued at over $1 billion.
In our latest @22Astronauts_ podcast, we discuss robotics infrastructure, raising venture capital, China's role in robotics, the future of software, and what it takes to build a company in Physical AI.
This episode is brought to you by @CoreWeave, providing the specialized cloud infrastructure powering physical AI and large-scale model training.
𝗧𝗜𝗠𝗘𝗦𝗧𝗔𝗠𝗣𝗦
00:00 Building the AI Robotics Engineer
03:25 Why customers buy solutions, not platforms
09:30 Scaling Eucalyptus and lessons from startups
17:30 Founding Alloy Robotics
22:45 Raising venture capital and startup storytelling
28:00 Building products customers actually want
36:30 China's robotics ecosystem and global supply chains
44:00 AI coding and the future of software
47:00 Build vs. buy in robotics
51:11 Why 99% robot reliability isn't enough
Full Episode:
🎥 YouTube: youtu.be/ufa9lJ01mF8
🎧 Spotify: open.spotify.com/episode/0k3…
Solution to a robotics search problem, (LeetCode 489, reported as a Google interview question):
A. Random bounce: given the same 200 tiles of driving, it misses 6 of the room's 12 inside corners
B. Spiral outward: breaks when it bumps the kitchen island
C. Follow the walls: never reaches the island, because the island touches no wall
D. Depth first search: every reachable tile clean, and it finishes on the tile it started from
This paper visualised: "Capture Point: A Step toward Humanoid Push Recovery" (Pratt, Carff, Drakunov & Goswami, Humanoids 2006).
- Model the robot as a linear inverted pendulum
- For any push, there's one spot on the floor where a single step stops it: x + v·√(h/g)
- Step short and it keeps falling, step too far and it tips forward
- Here: a MuJoCo G1 shoved back at 1.04 m/s steps 27 cm behind its hips and stops dead
"What really happened on Mars?" David Wilner's (Wind River) IEEE RTSS 1997 keynote, visualised: why Mars Pathfinder kept resetting itself a few days after landing.
- Low-priority weather task (ASI/MET) takes a mutex
- High-priority bus task (bc_dist) blocks on it
- Medium-priority tasks keep pre-empting ASI/MET, so it never lets go
- bc_sched sees bc_dist miss its 0.125 s cycle and resets the computer
The fix was one flag: priority inheritance.
Alloy Robotics retweeted
> be andromeda
> build humanoid robots
> not for a demo video
> for actual aged care homes
> fleet grows 4-5x in six weeks
> waves and waves of bugs
> appoint a "goalkeeper" to stop every shot
> week-long shift too exhausting
> plug in alloy
> when abi stops mid-session
> alloy answers before any engineer does
> fleet-wide audio bug: 2 days -> 10 min
> finds a bug across robots nobody knew was there
> l1 & l2 support: automated
> scale more deployments, faster
(ep 2 of real world robots)
"robotics isn't ready yet"
we're making a series showcasing robots working in the real world 🦾
ep1: halter, valued at $2B & backed by peter thiel
A MuJoCo sim of a Unitree G1 visualising this paper: "Zero-Moment Point: Thirty Five Years of its Life" (Vukobratović & Borovac, IJHR 2004).
At 6 kg: the zero-moment point slides toward the G1's toes, it leans back, and the point comes home.
At 16 kg: the point crosses the toe edge with the box at chest height. No foot left for the floor to push on.
This obstacle avoidance paper visualised: "Dynamic obstacle avoidance for quadrotors with event cameras" (Falanga, Kleber & Scaramuzza, Science Robotics 2020) teaches a drone to dodge a thrown ball.
- Swap the frame camera for an event camera
- Each pixel reports only when its own light changes
- Use event timing to tell the moving ball from the still room
- Detect the ball in 3.5 ms, with onboard sensing and compute only
They tested it on balls thrown from 3 m at 10 m/s.