CapraMorph
Desktop robot arm whose metal structure dissolves into a glowing orange wireframe toward the gripper, standing on a dark gridded floor.

CapraWorks · Robot learning

Policies come first. Morphologies later.

CapraMorph is a sim-first robotics research project. Control policies are developed and evaluated in MuJoCo against explicit exit criteria — hardware is the periodic materialisation of what survives the sim gates.

Phase 2 — picks it up, but cannot hold it: see log

Progress log Read the method hello@capramorph.com

01 · Method

Simulation is the primary instrument.

Development happens in simulation before anything is machined, printed or bought. The target platform is the SO-101, a low-cost desktop arm in the lineage of TheRobotStudio's SO-ARM100 — but the arm is the output of the process, not its starting point.

a

Train in sim

Policies are developed and evaluated in MuJoCo, where iteration is cheap and every run is measurable.

b

Gate on evidence

Each phase has explicit exit criteria, written down in advance. A phase either passes its gate or it does not.

c

Materialise periodically

Hardware builds happen only when the simulated evidence justifies them — the physical arm realises what already works.

Dozens of ghosted, translucent robot-arm trajectories; a single solid orange arm passes through a row of glowing rectangular gate frames.
Fig. 1 — Candidate trajectories against gate frames. One survives.

02 · Sim gates

Most candidates don't make it. That is the point.

Candidate policies are trained in simulation and judged against predefined exit criteria. Most are discarded. The few that pass a gate define what the next phase — and eventually the hardware — has to support.

The gates are written before the runs, so a pass means something and a failure is information rather than embarrassment.

Exploded blueprint view of a desktop robot arm: gripper, joints, servos and base separated vertically, annotated with orange schematic callouts.
Fig. 2 — SO-101 morphology, exploded.

03 · Morphology

Hardware is downstream of policy.

The target morphology is the SO-101, a low-cost desktop arm descended from TheRobotStudio's SO-ARM100 design. It is deliberately modest hardware: cheap enough to rebuild, well-understood enough to model.

Physical builds are periodic, not continuous. When a policy survives the sim gates, the arm is assembled to meet it.

04 · Current status

The arm picks things up.

Phase 0 (simulation substrate) passed 2026-08-14. Phase 1 — reaching a point — passed 2026-08-16 at 86.6%. Phase 2, as of 2026-08-18: the arm reaches for a box from a random start, grips it and lifts it 92.5% of the time over 80 trials, by chaining a pose-reaching policy with a grasp routine proven in advance on 4,258 certified positions. We previously published 85% and then 88.8% here. Both were wrong: our scoring counted a single 10 ms instant of contact as a successful pick, where our own standard requires the box be held for 27 of any 30 consecutive control steps. On the correct criterion this phase does not pass its own gate. The log has the full retraction. The progress log has the video, the numbers, and the day spent certifying a grasp that turned out to be the box held by the fingertips — caught by watching the footage, not by reading the telemetry.

Telemetry
Machine-readable logs from every training and eval run.
Video
Rendered rollouts of the learned policy, straight from the simulator.
Report
Deterministic eval numbers against each phase's pre-written exit criteria.

Learned policies in simulation; no hardware yet — hardware is the reward for passing gates. There is no product, no mailing list and nothing to sign up for. The log says exactly where things stand.

05 · Contact

Correspondence

Questions about the project are welcome. The working repository is private for now.