SO-100 vs LeKiwi: the same arm, plus a base you have to teach

The short answer

This is not two arms, it is one arm and the same arm on wheels. The manipulator on a LeKiwi is the SO-100 manipulator: five joints plus gripper, Feetech STS3215 servos, 7.4 V. So the question is not which picks things up better. It is whether you want the base, and whether you are ready to pay for it in demonstrations rather than euros.

Side by side

What the base adds to every episode

The action vector recorded on an SO-100 has six entries, one per actuator. On a LeKiwi the base adds its own commanded dimensions: three wheels, driven independently, which is what holonomic means here. Read the exact width off your recorded dataset, the authority on its own schema. The direction is not in doubt: the vector is wider, and no fixed arm ever had to predict those entries.

A policy has to produce every dimension it was trained on, at every step. It does not leave the wheels blank because the interesting part happens at the gripper. From the first episode, driving is part of what it imitates.

That changes what your episode budget has to cover. On a fixed arm the object sits in a few places and the camera never moves. On a mobile base, where the robot approaches from is a variable you are teaching. The fifty episodes GR00T N1.7 wants become fifty of driving and manipulating, and fifty from the same tile teach one approach path.

What the wheels do to teleoperation and recording

The recorder starts and stops the recording, it does not drive the robot. You teleoperate while it runs, which on a LeKiwi means driving the base and the arm inside one episode. Clumsy driving is written into the dataset as ground truth just as faithfully as clean driving.

Two rules follow. Keep an episode only when it was good: Enter ends a clean one, Ctrl-C discards a fumbled one. And settle the cameras first. A wrist camera moves with the arm on both robots, but a LeKiwi base camera moves with the machine, so the background shifts inside one episode.

What does not change

Which model copes best with a mobile base

There is no base specific policy here, so the question is which of the five copes with a wider action space. ACT has no pretraining and no language conditioning, so it learns your approach paths literally, and its 20 ms inference suits a driving robot. Pi0.5 at about 485 ms per action step needs care, because a base in motion makes a stale observation more expensive. SmolVLA sits between them at about 245 ms, useful from about thirty episodes.

A sane first week on a LeKiwi

1. 1
   Park the base and treat it as an SO-100

   Find the port, set the zero point, send it home. Everything you know about a fixed arm is valid at this stage.

   bash

   ay-robots devices
   ay-robots calibrate
   ay-robots home

2. 2
   Record a static baseline, then add driving

   With the base parked, record the task as on a fixed arm: that is your reference if a mobile policy later fails. Then record with the base moving, covering the approach directions you will need.

   bash

   ay-robots record --task "pick up the red cube and put it in the bin" \
     --dataset lekiwi-static --episodes 10 --cameras wrist,top
   
   # Then the same task with the base moving, from varied start positions.
   ay-robots record --task "pick up the red cube and put it in the bin" \
     --dataset lekiwi-mobile --episodes 30 --cameras wrist,top

3. 3
   Upload and train the cheap model first

   There is no ay-robots train command. Upload, then open Training, pick the dataset, choose SmolVLA, press Start. Agents call start_training instead.

   bash

   ay-robots datasets upload lekiwi-mobile --name "lekiwi cube pick"

4. 4
   Evaluate from marked starting positions

   Mark the floor, run the same positions in the same order, and write down where it failed: wrong approach, right approach with a bad grasp, or never near the object.

   bash

   ay-robots run --model smolvla \
     --checkpoint s3://ay-robots-runs/lekiwi-cube-pick/step-20000 \
     --task "pick up the red cube and put it in the bin"
   ay-robots stop

# Hosted MCP server: Streamable HTTP and JSON-RPC 2.0 at /api/mcp,
# protocol version 2025-06-18. Training tools take your API key.
curl -s https://www.ay-robots.com/api/mcp \
  -H "Authorization: Bearer $AY_ROBOTS_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "start_training",
      "arguments": {
        "dataset_id": "ds_lekiwi_cube",
        "policy_name": "lekiwi cube pick, smolvla",
        "model_id": "smolvla",
        "confirm_spend": true
      }
    }
  }'

Full against compatible, and debugging across the line

The SO-100 carries the full support level: it is the reference unit, and a symptom is reproduced on one before the page describing it goes up. The LeKiwi carries compatible. Its arm half is that same reference hardware, so the servo advice, the calibration procedure and the Feetech failure modes apply directly. Its base half is not.

So split the problem before you debug it. Park the base and run the task by hand. If the arm misbehaves standing still, it is an arm problem and every fix page applies as written. If the arm is fine and the approach is wrong, it is a data problem in the new dimensions.

Buy the LeKiwi if

• The task needs the robot to move: a second surface, a person to follow, a bench longer than the arm reach.
• Your fixed-arm workflow works and your failures are about reach, not grasping.
• You will record demonstrations covering several approach directions.

Stay on the SO-100 if

• You are learning the pipeline. Hours spent on driving are hours not spent on what decides whether a policy grasps.
• Your task fits on a table. A base adds dimensions, failure modes and a rail, and buys nothing if the object is in reach.
• Budget matters. Roughly 400 to 500 euro is three SO-100 arms: a teleoperation pair with a spare.

Where to go next

• /robots/lekiwi for the fact sheet, rails and base
• /robots/so-100 for the arm on top of it
• /train/smolvla-on-lekiwi for the cheapest first run
• /train/groot-n1-7-on-lekiwi once you have fifty clean episodes
• /fix/policy-only-works-in-one-setup when only one start position succeeds
• /fix/camera-not-detected when a base camera drops out
• /glossary/degrees-of-freedom for how the base widens the action vector
• /glossary/episode for what one demonstration contains
• /compare/so-100-vs-so-101 if you are still choosing the arm

Frequently asked questions