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.

1:30 a.m, 2016. A Roomba starts its nightly clean, finds fresh puppy poop and keeps going. Its wheels left 25-foot poop trails all over the house, and the owner's post passed 326,000 shares. A classic Roomba has no way to notice. Its bumper only feels things that push back, and its cliff sensors only look for drops. iRobot's fix was a front camera and a detector trained on hundreds of fake poo models.

This paper visualised: "Cloth Grasp Point Detection based on Multiple-View Geometric Cues with Application to Robotic Towel Folding" (Maitin-Shepard, Cusumano-Towner, Lei, Abbeel, ICRA 2010, UC Berkeley) 1. Hang the towel from one gripper so it falls into a simpler shape 2. Rotate it 12 degrees at a time in front of a stereo camera pair 3. Tell cut edges from folds by how sharply they curve across views 4. Grab two neighbouring corners, pull taut, fold twice

Bonsai Robotics head of AI John Macdonald on training for what the field rarely shows: - A flock of birds lands in front of the tractor and stays, and a model that doesn't know to push through is stuck - Bonsai World turns any farm in satellite imagery into a generative simulation of that farm - Add the rare cases (birds, wildlife, fallen trees) and train before the robot reaches the field

This control paper visualised: "A Controller Framework for Autonomous Drifting" (Hindiyeh & Gerdes, JDSMC 2014) holds a car at an unstable equilibrium on purpose. - In a drift the rear tyres sit on the edge of their friction circle - More throttle spends grip forward: the tail swings wide - Less throttle gives grip back sideways: the tail tucks in - The gas pedal steers the rear while the front wheels countersteer Car model: "Generic Sport Coupe Car" by MMC Works, CC BY 4.0

This imitation learning paper visualised: "DART: Noise Injection for Robust Imitation Learning" (Laskey, Lee, Fox, Dragan, Goldberg, CoRL 2017) makes robot demos worse on purpose. - Shove the human off course while they demonstrate - Size the shoves to the trained policy's error - The data fills with recoveries, with no expert labelling the robot's runs - Matches DAgger on MuJoCo Walker, Hopper, Half-Cheetah and Humanoid

This navigation paper visualised: "Navigating the Crowd: Non-linear MPC with Social Forces Dynamics for Human-Aware Robot Navigation" (Trepella et al., 2026) puts a model of the crowd inside the robot's planner. - Each person is pulled toward their goal and pushed away from anyone too close - Predict everyone 2 s ahead, including how they react to the robot - Take the path that stays out of everyone's personal space - Replan 20 times a second

This security paper visualised: "Towards Robust LiDAR-based Perception in Autonomous Driving" (Sun, Cao, Chen, Mao, USENIX Security 2020) puts a fake car in front of a self-driving car, then catches it by its missing shadow. - A photodiode syncs to the LiDAR; an infrared laser fakes echoes 5 to 8 m ahead - 60 fake points read as a car (a real one returns about 2,000): about 80% attack success - A real car blocks the laser, so nothing echoes from behind it. A fake one doesn't - CARLO checks behind every box for that shadow: attack success drops to 5.5% 3D cars: comrade1280 (Sketchfab), CC BY 4.0

This perception paper visualised: "Seeing Around Street Corners: Non-Line-of-Sight Detection and Tracking In-the-Wild Using Doppler Radar" (Scheiner et al., CVPR 2020) finds a cyclist hidden behind a building. - Radar waves are about 5 mm long, so a plastered wall acts like a mirror - Bounce off the facade across the street, round the corner, off the cyclist and back - Doppler keeps only what moves - Fold the hit back across the mirror wall to place the cyclist - Tracked to within about 10 cm 3D cars: comrade1280 (Sketchfab), CC BY 4.0

This safety paper visualised: "Control Barrier Function Based Quadratic Programs for Safety Critical Systems" (Ames, Xu, Grizzle, Tabuada, IEEE TAC 2017) puts a small optimisation between the AI's command and the motors. - h = blade-to-body distance minus a 2 cm margin - Solve min ||u − u_nom||² subject to ḣ ≥ −α·h - Approach speed is capped at α·h, zero at the margin - Safe commands pass through untouched - A knife driven at a man's back at 1.3 m/s stops 1.7 cm short

This paper visualised: "Vision-Based Navigation for the NASA Mars Helicopter" (Bayard, Conway, Brockers, Delaune, Matthies, Grip et al., AIAA SciTech 2019) - A downward camera tracks a few dozen pebbles as they slide between frames - A laser altimeter gives height, so the slide becomes speed - Over steep, featureless sand ripples, the features ran out - The speed estimate drifts, it lands moving sideways, hits a ripple, and all four blades snap Flight path reconstructed in a small sim; cause and outcome from NASA's accident investigation.

This SLAM(LOAM) paper visualised: "LOAM: Lidar Odometry and Mapping in Real-time" (Zhang & Singh, RSS 2014) on why a moving lidar bends straight walls. - Every point in a sweep fires at a different moment - Place them all from the first beam's pose and a turning robot sees the wall bow 0.58 m - Interpolate the pose at each point's own timestamp - Walls snap straight, and it makes the door

Antioch co-founder Harry Mellsop on the two things that robot simulations useful: 1. Fidelity (how much can you trust the sim?) 2. Scale (running thousands of them, not one).

This robotics manipulation paper that makes an arm play Jenga visualised: "See, feel, act: Hierarchical learning for complex manipulation skills with multisensory fusion" (Fazeli et al., Science Robotics 2019) - Push 1 mm at a time, logging wrist force and how far the block moves - 403 practice pushes sort into types: slides free, or force climbs and nothing moves - On a new block, compare the first few millimetres to each type - Clean extractions went from 56% to 88%.

Stereo cameras vs LiDAR. - Stereo finds the same patch in both images and reads depth from how far it shifts - A blank white panel has nothing to match: StereoSGBM on the rendered pair returns the trailer as speckle (median guess 102 m, truth 60 m) - Error grows with distance squared: with f = 1000 px and a 0.30 m baseline, half a pixel off at 60 m misses by 6 m - LiDAR just times the echo. 400 ns there and back is 60 m, and colour and contrast don't matter 3D models (CC BY 4.0): "Truck Trailer FREE" by NLM, "Semi Truck 1980" by Brian Trepanier, sedan model by BHP3D, all on Sketchfab.

SOLUTION EXPLAINED: NVIDIA Interview question reported on Glassdoor for autonomous driving roles. Why robots store rotation as four numbers, not three angles. - A drone tracks yaw, pitch and roll, one gimbal ring each - Pitch its nose to 90° and the yaw and roll axes line up: a degree of freedom vanishes from the maths (gimbal lock) - At 89°, yaw and roll swing 57x faster than the drone actually turns - A unit quaternion is a point on a 4D sphere, w² + x² + y² + z² = 1, with no dead spot, so every turn is a smooth path A numpy flight sim drives every ring with the real Euler-angle maths.

Knowing which item to pick is the easy part. Where to grab it is the problem. Closed-loop suction picking, shown in a 3D cell and in Nimble Robotics' real bins: - Isolate the ordered item from the clutter packed around it - Map candidate suction points on its visible surface - Score each point: can the cup land there and form a clean seal? - Lift, then check that exactly one object came up - Check fails? Move to a different contact point instead of stopping the cell

This paper visualised: "ClearGrasp: 3D Shape Estimation of Transparent Objects for Manipulation" (Sajjan et al., ICRA 2020) teaches a robot to see glass. - Depth light refracts through the glass, so the sensor reads the table behind it - Mask out every glass pixel, because that depth is lying - Predict surface normals and outlines from the colour image - Solve the missing depth back in, guided by the shape Two-finger picks of transparent objects: 12% to 72%.