Thomas Geijtenbeek

Emu model by @pashavanb.bsky.social, simulated using #depRL and #Hyfydy

Emu running simulation

This agent was optimized for maximum happiness

Jumping at 1x, 2x and 4x human strength

Someone asked if SCONE can do rotating treadmills, so I made this example

The model (32 dofs / 122 muscles) was trained using sconegym and deprl

Working on a new obstacle course for musculoskeletal models #Hyfydy

I'll be presenting our new predictive simulation model at ISB 2025 in Stockholm next week!

Parallel session 9F (Movement Generation and Prediction), Wednesday July 30 from 10:40 to 12:00.

Hope to meet you there!

These boxes are not moving. A mind-bending optical illusion by Japanese artist Jagarikin.

It’s the forward dynamics simulation that runs at 20x real-time

This model has 35 degrees-of-freedom, 120 musculotendon actuators and full contact geometry. Simulations run at around 20X real-time. No motion data was used.

Now featuring: a full upper-body biomechanical model that learned to reach (SCONE + Hyfydy)

Please consider adding me!

Please consider adding me here too!

Please consider adding me as well!

I simulate how animals move (see my feed for examples)

Muscle-Driven Predictive Physics Simulations of Quadrupedal Locomotion in the Horse Synopsis. Musculoskeletal simulations can provide insights into the underlying mechanisms that govern animal locomotion. In this study, we describe the dev

Model developed by @pashavanb.bsky.social

academic.oup.com/icb/article/...

My favourite local minimum

The scapula is kept in place through spring forces between points on the scapula and an ellipsoid on the thorax. Springs that push the points outwards have a higher stiffness than springs that pull the points towards the ellipsoid.

The muscle activations are optimized for the hand to reach the red sphere targets; the scapular motion emerges independently.

Our biomechanical shoulder model now has full contact geometry, so the arm must navigate around the torso to reach behind...

See the link in my other reply

D-Flow 🫶

Natural Walking RL Preprint

More information: sites.google.com/view/natural...

This uses a 30-dof model with 122 compliant musculotendon units, simulated using Hyfydy and trained using depRL. No motion data was used for training.

Now that Bluesky has native video support, I figured it's time for a repost: Teaching a biomechanical model to run using reinforcement learning.

New results coming soon! #BiomechSky

Happy Halloween! (Sound on 🎶)

No skeletons were harmed in the making of this video 😇

Real-time interaction with a musculoskeletal model 💀

Trying to teach my new model to run using reinforcement learning Created using the Hyfydy simulation engine (https://hyfydy.com).Trained using depRL (https://sites.google.com/view/naturalwalkingrl).No motion data was used ...

Trying to teach my new model to run using reinforcement learning 😅
youtu.be/cnoJXmc5-gs