Jack Dubie
-
Eight robots running 97.6% autonomously
Eight robots packing at a customer site, running 97.6% autonomously. They have packed over 45,000 bags for our customer.
-
Teaching Gemma 1B Wordle: SFT for the rules, RL for the strategy
Pretrained Gemma 3 1B goes from 0% to 80% at Wordle with SFT, and to 85% with RL on top.
-
RL environments that hold up when the model tries to cheat
Sandboxed coding environments, three graders, and what GRPO on Gemma 3 1B learns when the grader can be fooled.
-
Teaching a Unitree G1 not to trip
PPO in Isaac Lab takes a Unitree G1 from falling on 1 in 18 runs over 1-inch clutter to 0 falls with a natural 0.9 Hz stride, and shows what sim-to-real realism costs.
-
Coding agents stacking blocks with a robot arm
Six coding agents each got 20 minutes and a 300 Hz socket to a simulated YAM arm; Opus 5.5 and GPT-6.1 Sol both stacked all eight blocks.
-
Teaching Gemma 1B made-up puzzles with reinforcement learning
GRPO takes Gemma 3 1B from 33–59% to 96–100% on made-up puzzles in 45 minutes on one TPU chip.
-
A PostgreSQL 16 server, wire protocol and all, where every persisted file is empty
Six coding agents had 20 minutes to write a PostgreSQL 16-compatible server whose saved files must all be empty; Opus 5.5 (0.655) and GPT-6.1 Sol (0.649) led, and nothing came close to saturating it.
-
Black-box database: rebuild a server you can only probe
Six coding agents had 20 minutes to reverse-engineer a database they could only probe over TCP, then rebuild it byte for byte; GPT-6.1 Sol scored 0.941, and memorising PostgreSQL is worth almost nothing.
-
Fine-tuning π0 with zero robot data
Policies fine-tuned only on exoskeleton demonstrations, retargeted to the robot with offline iterative learning control, with zero robot-collected samples: a cube in November, a pile of bags in December.
-
Teaching Gemma 1B to count letters
Fine-tuning Gemma 3 1B on synthetic examples takes letter counting from 20% to 84%.
-
Can a coding agent design a car and drive it?
Seven coding agents got 30 minutes and 1,000 attempts to design and drive a 2D car over rough terrain, through a socket; GPT-6.1 Sol, Opus 5.5 and GPT-6.1 Terra went about 2.4 km, Haiku 5.5 about 0.4 km.
-
Meta Quest 3S UMI implementation
Replacing an $8,000 exoskeleton with a $500 Meta Quest 3S for data collection: 20% lower tracking error, and a policy picking at 4.4 seconds per pick with zero robot data.
-
How good are agents at robotic control?
An agent benchmark where models get 8 minutes and a socket to a simulated YAM arm whose dynamics are hidden, and must make it follow 14 trajectories. Sol (gpt-6.1-sol) reached 0.10 mm, better than a controller that knows the true dynamics; the Claude models stopped early.
-
Gemma 1B won’t fix its own code
Gemma 3 1B resubmits the same failing code 95% of the time; starting over works far better than asking it to fix (34% vs 1%).
-
Physical Intelligence featured our robot
Physical Intelligence featured Frontier Machines in their Pi API beta launch. Their base model is really impressive.
-
Can a coding agent CAD a robot finger from pictures?
Five coding agents got five pictures of a robot gripper finger and 20 minutes to rebuild it in CAD; Opus and Haiku matched a careful hand-built model, and most quit within six minutes.
-
Removing chunk seams from π0.5 with training-time RTC
Implementing training-time realtime action chunking in π0.5: a small change to training and inference that made rollouts smooth, stateful and about 99% autonomous on 14 hours of teleop data.
-
Reproducing DSRL: latent-space RL on pi0, from sim to a real robot
Reproducing DSRL, latent-space RL on pi0: it took a sim cube pick from 50% to about 95% and a LIBERO task from 30% to 100%, then lifted a real arm's candy-bar toss from 16% to 41%.
-
Make the tests faster, without breaking them
An RL environment that rewards coding agents to speed up a slow client-server test suite only if it still catches 24 hidden bugs, where Gemini 4 Argon reached 40.8x and Opus 10.3x against a hand-written 78.7x. Grok 4.7 scored 0: its suite was 23.5x faster and caught every bug, but it loosened a rate limit to get there.
-
Cleaning up toys from 7,000 miles away
A home robot with under $1k of parts, leader-follower arms and an internet teleop link, picking up toys while I drove it from 7,000 miles away.
-
Find every duplicate file, as fast as you can
Six coding agents had 15 minutes, six times each, to write the fastest correct duplicate-file finder for 23 hidden trap trees. Opus was fastest on average; Gemini matched it when it finished.
-
Real2sim with a coding agent
A coding agent rebuilt a bottle case-packing cell in MuJoCo from the robot's own cameras in about an hour and a half, iterating until its renders matched the real footage.
-
Which school of ethics does your AI model follow?
Haiku is a deontologist, Opus follows the Doctrine of Double Effect, and the other five AI models are utilitarians.
-
Remote teleoperation at 85 ms glass-to-glass
A remote teleop stack for dexterous manipulation on Meta Quest 3S, I2RT YAM Pro arms and an NVIDIA Jetson Orin with GMSL2 cameras: 85 ms glass-to-glass with operators 2,000 miles away.
-
Cutting tracking error on compliant arms and an exoskeleton
Reducing YAM position tracking error 87% (30 mm to 4 mm) so position-control policies could run on compliant arms, calibrating an exoskeleton against OptiTrack (40 mm to 6 mm), and comparing UMI tracking options end to end.
-
Doubling throughput with fingers and jigs
On a bimanual parallel-gripper robot, task-specific fingers and 3D-printed jigs doubled task throughput and gave the wrist camera a much clearer view.
-
Sensor fusion from scratch: tuning an IMU + camera + LiDAR Kalman filter in simulation
A from-scratch error-state Kalman filter for IMU + camera + LiDAR, tuned in simulation over ~1,000 Monte Carlo runs: Q/R tuning with NEES and NIS, outlier gating and its lock-out failure, latency, camera time offset, online extrinsic calibration, and sensor dropout.