Alexander Koch @alexkoch_ai Twitter profile

Last Seen Profiles

@dwdoii

@RonsEmma

@biomejs

@DrJulieHarris

@Shadmirable

@siema161

@niiconn

@Sarneics_VRC

@femboygoetia

@_DJDON

@_tsukuyxmi

@SavvyAuntie

@Havas

@yonkoujean

@HylandIan

@Naririn79153615

@iampriya06

@MaggieLombardi2

@LaboralXxi

@mobilerecell

@M_alhilal_97

@nonstop_live

@SubZeroMT96

@SadeNotSadiee

@YilmazGundogdi

@XAedroHORJoMXId

@opendtect

@zyon_mccollum

@ixlino

@NikilisRBX

@abubake24596102

@RitaRys1

@keisemi

@PSMMAzadmanesh

@mura_buta

@ueno_wan

Alexander Koch

@alexkoch_ai

2 months

Folding clothes with $250 robot arms. I've added another motor to improve mobility and extend the reach. The CAD files and the code are public at: (video at 2x speed)

129

389

3K

Alexander Koch

@alexkoch_ai

3 months

Early results from my AI training runs. I've trained my $200 robot arm on a simple picking task using imitation learning. It has learned to control the robot arm using only camera images and joint states.

94

286

2K

Alexander Koch

@alexkoch_ai

16 days

Excited to announce Tau Robotics ( @taurobots ). We are building a general AI for robots. We start by building millions of robot arms that learn in the real world. In the video, two robot arms are fully autonomous and controlled by a single neural network conditioned on different…

123

318

2K

Alexander Koch

@alexkoch_ai

3 months

Added more arms. The entire setup is now about $600.

65

91

1K

Alexander Koch

@alexkoch_ai

2 months

Open-source robot arm for about $200. It uses five Dynamixel servo motors and weighs slightly over 100g (without the base). The design for a leader arm is included as well so that you can teleoperate it effectively. (Video is at 1x speed)

28

130

872

Alexander Koch

@alexkoch_ai

2 months

$1000 robot farm

32

49

564

Alexander Koch

@alexkoch_ai

3 months

Finished building the first version of my low-cost teleoperation setup. It's less than $200 per robot arm.

31

36

556

Alexander Koch

@alexkoch_ai

2 months

Trained a simple world model for my robot arm. It predicts the future over 20000 times faster than real time on a single @nvidia RTX 3090 GPU (128 batch -> 160x faster each).

14

57

532

Alexander Koch

@alexkoch_ai

1 month

The $250 open-source robot arm is on the front page of Hacker News currently. Nice discussion at

14

34

380

Alexander Koch

@alexkoch_ai

30 days

There is now a MuJoCo simulation for the low-cost robot arm. This can be used for testing and reinforcement learning. Contributed by @arbwes .

11

57

387

Alexander Koch

@alexkoch_ai

2 months

Upgraded my robot arm to use a Raspberry Pi. Now it only requires a single 12v power supply and a data cable. The Raspberry Pi will send the sensor data via ethernet to a central GPU server to compute new actions. This server should be able to control 100 robot arms in parallel.

15

25

335

Alexander Koch

@alexkoch_ai

3 months

The video shows a policy that was trained on two instructions. One instruction to pick up the block and one to put it down. The entire training dataset contains about 400s of teleoperated trajectories.

3

6

128

Alexander Koch

@alexkoch_ai

3 months

The video itself was filmed by another robot arm

3

7

125

Alexander Koch

@alexkoch_ai

3 months

The learning algorithm is based on ACT from @tonyzzhao . However, it is modified to use language instructions to learn multiple tasks with a single neural network.

2

3

92

Alexander Koch

@alexkoch_ai

3 months

@gazorp5 I've designed and built it myself. The two servos in the base are Dynamixel XL430-W250-T. The other three servos are Dynamixel XL330-M288-T.

Alexander Koch

@alexkoch_ai

3 months

Finished building the first version of my low-cost teleoperation setup. It's less than $200 per robot arm.

31

36

556

2

8

86

Alexander Koch

@alexkoch_ai

2 months

The assembly takes about 20 minutes by an expert. I've uploaded the full assembly video on YouTube (). The designs are optimized for 3D printing. Only one part needs supports.

2

5

66

Alexander Koch

@alexkoch_ai

2 months

It predicts future camera images and joint angles given the control input from the leader arm. This learned simulation will later be used to train an AI to control the robot arm. My implementation is based on the Dreamer algorithm by @danijarh .

2

1

55

Alexander Koch

@alexkoch_ai

2 months

@gazorp5 Yes, but I’m working on the autonomous execution

4

0

46

Alexander Koch

@alexkoch_ai

2 months

It has a payload of a few hundred grams and a reach of slightly over 20cm. This Apple charger weighs just over 300g.

1

4

39

Alexander Koch

@alexkoch_ai

3 months

@faraz_r_khan @tonyzzhao It is conditioned on language instructions. I'm alternating two instructions in the video so that repeatedly picks up the block and puts it down. The entire input is the camera image, the joint states (position + velocity), and the language instruction.

1

39

Alexander Koch

@alexkoch_ai

2 months

It is well suited for robot learning. The low weight and torque make the arm inherently safe.

Alexander Koch

@alexkoch_ai

3 months

Early results from my AI training runs. I've trained my $200 robot arm on a simple picking task using imitation learning. It has learned to control the robot arm using only camera images and joint states.

94

286

2K

2

3

39

Alexander Koch

@alexkoch_ai

2 months

The design is inspired by the ALOHA project from @tonyzzhao and the lower-cost version GELLO by @philippswu .

0

2

32

Alexander Koch

@alexkoch_ai

2 months

@0xreklaw Maybe they can already. I haven’t tried it yet

2

0

22

Alexander Koch

@alexkoch_ai

3 months

@MatthewBerman Yes. Although the main inspiration was the ALOHA paper by @tonyzzhao that preceded mobile ALOHA. Further inspiration was the GELLO project by @philippswu who proposed a cheaper version of the ALOHA teleoperation system.

0

17

Alexander Koch

@alexkoch_ai

2 months

@HououinTyouma That's easy to do. But you'll probably have to add one more axis.

Alexander Koch

@alexkoch_ai

3 months

Added more arms. The entire setup is now about $600.

65

91

1K

0

3

13

Alexander Koch

@alexkoch_ai

2 months

@ThatFakeGuyDan $1000 for four arms

1

0

8

Alexander Koch

@alexkoch_ai

3 months

@BurnsworthRyan I am using Dynamixel XL330-M288-T except for the two shoulder joints in the teleoperated arm. Those are Dynamixel XL430-W250-T which have more torque.

0

8

Alexander Koch

@alexkoch_ai

3 months

@interwound About 25ms in software. Most of the delay in teleoperation is due to the inertia of the follower robot arm. However, my robot arm is very lightweight so the delay is short.

0

5

Alexander Koch

@alexkoch_ai

2 months

@drew_carson_ai A few hundred grams

2

0

7

Alexander Koch

@alexkoch_ai

1 month

@thejamesvance @gazorp5 I'll use a learned simulator. The low-cost robot arms make it possible to build a massive robot farm in the real world and collect a dataset to learn a world model.

Alexander Koch

@alexkoch_ai

2 months

Trained a simple world model for my robot arm. It predicts the future over 20000 times faster than real time on a single @nvidia RTX 3090 GPU (128 batch -> 160x faster each).

14

57

532

1

0

6

Alexander Koch

@alexkoch_ai

15 days

@seclink @taurobots The arm design is open source:

GitHub - AlexanderKoch-Koch/low_cost_robot

Contribute to AlexanderKoch-Koch/low_cost_robot development by creating an account on GitHub.

github.com

0

5

Alexander Koch

@alexkoch_ai

2 months

@RousslanDossa @nvidia I've previously implemented it in Jax but switched to PyTorch. The Dreamer algorithm compresses the observations into a small latent space and predicts only in this compressed space without reconstructing the images. This makes it very fast.

0

4

Alexander Koch

@alexkoch_ai

3 months

@iiJDSii Yes, the handle and trigger designs on the leader arm were created by @philippswu . I designed all the other 3D-printed components myself.

1

0

4

Alexander Koch

@alexkoch_ai

3 months

@RyanMcIntosh33 If you have some money, you can buy two arms from trossenrobotics ()

Research Manipulator Arms | Trossen Robotics

Explore our renowned X-Series Robotic Research Manipulator Arms. Up to 750g of payload capacity and up to 6 Degrees of Freedom (DoF)

www.trossenrobotics.com

1

0

4

Alexander Koch

@alexkoch_ai

3 months

@natfriedman sure, I've sent you a dm

1

0

4

Alexander Koch

@alexkoch_ai

15 days

@marshallk @taurobots yes

0

3

Alexander Koch

@alexkoch_ai

3 months

@zahi_kakish I've programmed it myself using PyQt

0

2

Alexander Koch

@alexkoch_ai

3 months

@danbri Initially, I designed the robot to have a camera on the wrist and a camera on each shoulder. However, the wrist camera adds some weight and is quite blurry due to the movement. And the shoulder cameras add some complexity since you need two of them.

1

0

2

Alexander Koch

@alexkoch_ai

2 months

@YeojMars I've used the Creality Ender 3 V3 SE and the Bambu Lab P1P. The P1P is better but also more expensive.

0

2

Alexander Koch

@alexkoch_ai

2 months

@kenny__shaw The latency is almost 10x lower when connected to a macbook. The U2D2 works fine with Linux. Also see here:

U2D2 high Latency

If I use the U2D2 to communicate with my Dynamixel XL330-M288 from my macbook (macos) I always have a latency of over 15ms for reading the current position or setting the goal position for example. I...

forum.robotis.com

0

1

Alexander Koch

@alexkoch_ai

1 month

@lukas_m_ziegler

Alexander Koch

@alexkoch_ai

3 months

The learning algorithm is based on ACT from @tonyzzhao . However, it is modified to use language instructions to learn multiple tasks with a single neural network.

2

3

92

1

0

1