Case study · Robotics · Aether for robotics

Strider Robotics

Took a humanoid manipulation policy from lab demo to factory line in one quarter.

Headline

1 quarter

sim-trained policy → production line

4 → 1

bespoke sim tools retired

1

world model, two embodiments

First deploy

transfer held, no per-task re-tuning

Background

Strider builds bipedal humanoids for warehouse manipulation. Their stack was the usual embodied-AI scaffolding rebuilt per task — a MuJoCo rig for prototyping, an Isaac pipeline for scale, a teleop data farm, a homegrown RL trainer, and a sim-to-real calibration step that exactly one engineer owned. Every new task meant standing the scaffolding back up.

The challenge

Their bottleneck was never policy architecture — it was the world the policies trained in. Teleop data collection was the long pole, and sim-to-real transfer was fragile and re-tuned by hand for every task. A policy that looked finished in low-fidelity sim could still fail on contact-rich work on real hardware, and the team had no honest signal for which was which until the robot tried it.

What we did

Ported three production tasks — bin picking, two-handed handoff and cable insertion — onto SimOS as a single shared world model.
Retargeted roughly 200 teleop demonstrations across two humanoid embodiments; SimOS augmented them under physically-plausible domain randomization into a full training distribution.
Rendered synthetic RGB-D and tactile in the loop, consistent with the same contact physics the controller saw — no separate perception-sim pipeline.
Ran system identification from a handful of real rollouts to calibrate the reality gap, with the model reporting per-task where transfer was trusted.

In their words

“The model told us which task wasn't ready before the robot did. That's the part we didn't have before.”

— Head of Learning, Strider Robotics

The outcome

Bin-picking and handoff policies transferred on the first hardware deploy at the predicted success rate, with no per-task transfer re-tuning.
Cable insertion was flagged out-of-domain at real-time fidelity and held back for high-fidelity contact training rather than shipped on a false-confident rollout.
Teleop data collection left the critical path — the same ~200 demonstrations covered all three tasks through cross-embodiment retargeting.
Four bespoke tools — the MuJoCo rig, the Isaac pipeline, the homegrown trainer and the transfer-calibration script — collapsed onto one world model.

One detail worth calling out

The honest win was the cable-insertion call. The old pipeline would have shipped a policy that looked good in a low-fidelity sim and failed on contact-rich insertion on hardware. SimOS reported the reality gap as too wide at real-time fidelity and required high-fidelity contact before it would certify transfer — the team caught a would-be field failure before it ever left the lab.