Rig

Everything on this site ran on this table.

• ≈ €1,350owned hardware (paid; 3 small parts not priced yet)
• $45rented GPU and API for the 22 register tests
• simmost common setup behind the papers' own claims (9 of 22; 10 on rigs over $5k)

Parts

• 

  SO-101 follower arm — the arm every trial ran on

  • 6 DoF, Feetech STS3215 servos
  • position control, no force or torque sensing
  • G-clamped to the back edge of the table

  €350 paid (pack price, with the leader arm and the wrist camera)

  Limits: short reach and no contact sensing: a grasp is only known to have worked by looking

  all experiments

• 

  SO-101 leader arm — teleoperation, for recording demonstrations

  • same servos, passive gripper handle
  • packed away, out only for recording sessions

  included with the SO-101 follower arm

  Limits: setting up teleop takes time, so new data comes in batches

  data

• NVIDIA Jetson Orin NX 16 GB (Super dev kit) — the only computer on the robot

  • JetPack 6.2 / L4T R36.4.3, CUDA 12.6
  • MAXN_SUPER + jetson_clocks for benchmarks, 25 W otherwise
  • headless

  €950 paid

  Limits: no desktop GPU: models that need flash-attention kernels without an Orin build (REKI) run on a rented GPU

  all experiments

• Logitech C920 (top camera) — the judge's view and the policy's main view

  • 640×480 at 30 fps, OpenCV
  • oblique mount; no overhead view possible here

  €50 paid

  Limits: perspective view lit by a window: colour masks depend on the time of day

  all experiments

• icSpring USB camera (wrist) — close view of the gripper

  • 640×480 at 30 fps, OpenCV

  included with the SO-101 follower arm

  Limits: narrow field of view; often sees only the bowl it is holding

  all experiments

• Green cutting mat — work surface

  • fixed colour background for the colour masks

  price to come

  Limits: one surface, one colour: no texture or lighting variation unless added on purpose

  all experiments

• Folding table — holds the station

  • station stays installed; leader packed away

  price to come

  Limits: moves on contact, so hard pushes shift the camera geometry

  all experiments

• Clamp, USB hub, power supplies, cables — everything else

  • one G-clamp, one USB hub, servo and Orin power supplies

  price to come

  Limits: a loose USB link drops servo packets (seen on motors 4 and 5)

  all experiments

Rented GPUs (vast.ai) — training, and serving models the Orin cannot run. RTX 3090 / 4090, A6000, RTX PRO 5000 class, by the hour · destroyed after each run. Total to date: $45 (scoreboard).

What this rig cannot test

• SO-101 follower arm: short reach and no contact sensing: a grasp is only known to have worked by looking
• SO-101 leader arm: setting up teleop takes time, so new data comes in batches
• NVIDIA Jetson Orin NX 16 GB (Super dev kit): no desktop GPU: models that need flash-attention kernels without an Orin build (REKI) run on a rented GPU
• Logitech C920 (top camera): perspective view lit by a window: colour masks depend on the time of day
• icSpring USB camera (wrist): narrow field of view; often sees only the bowl it is holding
• Green cutting mat: one surface, one colour: no texture or lighting variation unless added on purpose
• Folding table: moves on contact, so hard pushes shift the camera geometry
• Clamp, USB hub, power supplies, cables: a loose USB link drops servo packets (seen on motors 4 and 5)

No force, one arm, no scale; every write-up says which of these applies to it.

Not owned yet

• Second SO-101 (assembled) — a second arm for two-arm tasks status: not owned
• reBot B601-DM — upgrade path if reach becomes the limit status: not owned
• Desktop GPU, Jetson Thor or DGX Spark — considered for on-site training and for models with no Orin build status: not owned

Photo credits

• SO-101 follower arm: source, Apache-2.0 (TheRobotStudio/SO-ARM100)
• SO-101 leader arm: source, Apache-2.0 (TheRobotStudio/SO-ARM100)

Rig profile and calibration shape: paper-vs-hardware/rig/. Prices marked paid are what I paid, in euros; the source of every price is in data/rig.yaml, and none is a guess.

Reproducibility

Every post's numbers were recomputed cold on 2026-10-31, from a fresh clone of the public repository, the public Hub with no token and the published container, on this Orin NX at 25 W: bars recomputed from the published rows on 29 of 29 posts, gates rerun on the Orin for 6, none contradicted. The table, the versions and the tolerance rule: paper-vs-hardware rig/REPRODUCIBILITY.md.