A robot arm halts over a part it has not seen, an operator's gloved hand guides it through the motion once, and the arm continues alone

MA Reflex / The adaptive runtime for robots

Show it once. It does the job.

MA Reflex gives a machine one-shot learning, lasting memory and a safety gate on every action, on top of the policy it already runs. On a CPU, at the edge, offline.

1

Demonstration per job

270 / 270

Placements on a new robot, in simulation

1.68 ms

Per gated action, on a CPU

01 / Why a runtime

Why robots need a runtime.

A robot is expensive to teach. A new task means data collection, training time and an integration project. A new machine means doing much of it again. MA Reflex removes most of that loop, and it keeps working when the scene changes.

Your hardware, sensors, perception and models stay as they are. MA Reflex is the runtime layer on top of them.

The autonomy system
A quadruped robot climbing stairs at an industrial plant

03 / Day one

What changes on day one.

A robot on an assembly bench being fitted and tested
  1. 01

    It sets itself up.

    Connect the machine and MA Reflex works out its own body. In simulation, on simulated models of three commercial arms, that took under five minutes.

  2. 02

    The job is shown once.

    Where the tool goes, when the gripper closes, where things sit. One demonstration. Nothing is trained.

  3. 03

    Every action is gated.

    From the first move, confident actions commit, uncertain ones defer and unsafe ones are refused.

  4. 04

    Every action is on record.

    Each lesson and each action lands in a signed, tamper-evident record.

04 / From plug-in to production

From plug-in to production.

  1. 01

    Plug in

    MA Reflex works out the machine's own body.

  2. 02

    Show once

    Demonstrate the job a single time. Nothing is trained.

  3. 03

    Practice

    It rehearses where it was shown, accounts for every phase and finds the grasp that holds.

  4. 04

    Run

    It does the job at new layouts and on other robots, with every action passing the safety gate.

05 / Self-setup

A robot nobody described sets itself up.

No description file to write, no calibration day. Measured in simulation, on simulated models of three commercial arms, it took 193 to 296 seconds, inside a five-minute budget.

A robot arm on a test stand moving one joint at a time through a slow self-check

06 / One-shot learning

One demonstration. Any layout. Another robot.

In simulation, a job shown once on one robot was carried out by a different robot at 30 new layouts. 270 of 270 placed, zero retries, median placement error 0.57 to 2.45 mm. With noise added to the camera frames, still 270 of 270.

A collaborative robot arm moving a circuit board from a stack to a tray

270 / 270

Placed on another robot, in simulation

0

Retries

0.57–2.45 mm

Median placement error

07 / Beside your policy

Makes great policies better.

MA Reflex shares the job with your policy phase by phase. Across forty tasks on a simulation benchmark, success with objects moved rose from 66% to 79%, and from 39% to 65% with the object 21 cm away. The policy was asked 13.4 times per episode instead of 32.8.

A gantry robot drilling an aircraft wing panel

79%

Success with objects moved, from 66%

65%

With the object 21 cm away, from 39%

13.4

Policy calls per episode, from 32.8

Forty tasks on a simulation benchmark.

08 / Safety gate

Commit. Defer. Refuse.

Every action passes a gate. Confident actions commit. Uncertain ones defer to the policy or the operator. Unsafe ones are refused or held inside the keep-out zones you set, and the system fails closed. A corrupted sensor reading is refused, not acted on. Limits can be added or released at runtime.

Safety and governance
A robot working beside a marked keep-out zone, the gate state changing as it nears the edge

09 / Lasting memory

Skills that never overwrite each other.

Teaching the tenth skill leaves the first nine exactly as they were. MA Reflex recalls the right skill from context, combines known skills for new goals and chains them into monitored multi-step jobs. It remembers places and what worked there. Remove a skill and nothing else moves.

Close detail of a robot gripper closing on a partClose detail of a robot gripper closing on a part

10 / Edge native

Runs on a CPU. Offline. Air-gapped.

No GPU, no cloud dependency, no data leaving the site. A gated action takes 1.68 ms and the loop clears 20 Hz on a CPU. Every lesson and every action lands in a signed, tamper-evident record, and field experience can be exported as a signed training set.

A compact embedded computer on a robot chassis, cables and a status lightA compact embedded computer on a robot chassis, cables and a status light

1.68 ms

Per gated action, on a CPU

20 Hz

Control loop cleared, on a CPU

11 / Autonomous skill acquisition

A new skill, acquired on its own.

With no policy, no demonstration and no reward, MA Reflex can acquire a new skill on its own, safely under the gate, and keep it.

A robot arm working alone in a quiet cellA robot arm working alone in a quiet cell

12 / For operators

What operators get.

A driverless pallet truck setting a pallet down in a floor bay and reversing out
  1. 01

    Teach by showing.

    Show the job once. Nothing is trained, and the job carries to new layouts.

  2. 02

    Limits they control.

    Keep-out zones and limits added or released while the machine works, and held on every action.

  3. 03

    A clear answer on every action.

    Commit, defer or refuse. When it defers, the decision comes to them.

  4. 04

    A record they can check.

    Every lesson and every action, signed and tamper-evident, verifiable while it runs.

13 / For engineers

What engineers get.

A sensor head turning on a machine at dusk, the surroundings reflected in its glass
  1. 01

    Connectors for the stacks you run.

    Hugging Face LeRobot, SO-101, NVIDIA Isaac, ROS 2 and MoveIt 2, and on-device vision.

  2. 02

    A certification battery.

    Know before deployment whether an adapter passes.

  3. 03

    A Python SDK and HTTP API.

    Every capability, from teaching a skill to reading the record.

  4. 04

    A signed training set.

    Field experience exported to make the base policy stronger.

Integrations

14 / Where it runs

On the machine. Not in somebody's cloud.

MA Reflex runs on the machine or on an edge computer beside it, on a CPU. It needs no network, so it works offline and air-gapped, and nothing leaves the site unless you send it.

  • On the machine or an edge computer beside it
  • CPU only, no GPU
  • Offline and air-gapped
  • A single edge node or a whole fleet
  • On your premises
A robot arm placing a circuit board in a tray, a driverless pallet truck, a small delivery robot on a sidewalk, aircraft in formation, a drone lowering a parcel on a tether and a humanoid robot carrying a tote

15 / What it costs to run

No GPU to buy. No training run to pay for.

MA Reflex runs on the CPU a machine already carries. A new job takes one demonstration, not a training run. And every call it takes off the policy is compute you no longer spend.

A small rounded delivery robot driving along a city sidewalk

0

GPUs required

0

Training runs for a new job

2.4×

Fewer policy calls, on a simulation benchmark

1.68 ms

Per gated action, on a CPU

16 / How a pilot runs

How a pilot runs.

A factory floor at shift change
  1. 01

    Start in simulation

    The whole loop runs on a laptop, with no hardware and no cloud.

  2. 02

    One machine, one job

    On your floor. MA Reflex sets the machine up and is shown the job once.

  3. 03

    Measure it

    Success, policy calls and every gated action, all on record.

  4. 04

    Roll it out

    For your machines, your lines or your fleet.

Partners

17 / Proof

One demonstration, measured across robots and tasks.

Measured in simulation, on simulated models of commercial arms and on a forty-task simulation benchmark.

No.TestResult
01A robot nobody described sets itself up193 to 296 s, inside a 300 s budget
02A job shown once on one robot, done by another at 30 new layouts270 of 270 placed, 0 retries, 0.57 to 2.45 mm median error
03Forty tasks, scenes as set96% with MA Reflex beside the policy, matching the policy at its best
04The same tasks with the object or destination moved79% with MA Reflex, 66% for the policy alone
05The same tasks with the object 21 cm away65% with MA Reflex, 39% for the policy alone
06Calls to the policy per episode13.4 with MA Reflex, 32.8 for the policy alone
07MA Reflex alone: no policy, no GPU, no training74%, 63% and 57% on the same three settings

Measured in simulation. Rows 01 and 02 ran on simulated models of commercial robot arms. Rows 03 to 07 come from a simulation benchmark of forty tasks. Scene by scene, the shared job succeeded where the policy alone failed 129 times, against 46 the other way round.

Every result

18 / Questions

Questions.

  • Does MA Reflex replace my policy?

    No. It runs on top of it. MA Reflex takes the phases it knows and the policy keeps the rest.

  • What if we do not run a policy yet?

    MA Reflex can be delivered together with a vision-language-action policy, integrated as one complete system. Open policies are supported through the Hugging Face LeRobot integration.

    Ways to work with us
  • Do I need a GPU?

    No. It runs on a CPU, offline and air-gapped if you need it to.

  • Which machines does it work with?

    Any machine it can read state from and command. MA Reflex is coming to every machine that moves.

  • Can we start before touching hardware?

    Yes. The whole loop runs in simulation on a laptop, with no hardware and no cloud.

  • What happens when it is not sure?

    It defers. The decision goes to your policy or to the operator, and the deferral is on record.

  • What if a sensor reading is wrong?

    A corrupted or out-of-range reading is refused, not acted on, and the system fails closed.

  • How does it keep a robot safe?

    Every action passes a gate that commits, defers or refuses, inside the keep-out zones you set, and the system fails closed. It runs alongside your hardware e-stop and safety controller.

  • Does it need a network?

    No. It runs on the machine or beside it, offline, on a single edge node or across a fleet.

  • Can it learn with no policy at all?

    Yes. It can acquire a skill on its own, exploring safely under the gate.

  • How do we get it?

    Machine Autonomy offers MA Reflex to robot makers, integrators and operators, on its own or with a policy as one complete system. It starts with a pilot on one machine and one job.

    Ways to work with us
A robot cell at rest, lights lowA robot cell at rest, lights low

Give your machines a reflex.

Bring one machine and one job. We will show it once and let MA Reflex take it from there.