Here at #EWRL: demonstration of autonomous tomato harvesting by polybot.eu
#Robot #Harvesting #Learning

Fascinating talk by Amy Zhang
on Proto Successor Measures at #EWRL 2025.
euro-workshop-on-reinforcement-learning.github.io/ewrl18/progr...
#reinforcementlearning #robotics #machinelearning

Sergey Levine was just presenting in the Exploration in AI @ #ICML2025 and promoted that exploration needs to be grounded, and that VLMs are a good source ;-) Check our paper below
👇

Getting from Tübingen participating in the Huggingface LeRobot hackathon!
#Huggingface #Robotarm #hackathon #PhysicalAI #TuebingenAI

huggingface.co/spaces/LeRob...

Excited to be at the opening of the BITS: Bionic Intelligence Tübingen Stuttgart.
www.bionic-intelligence.org
Intro by Syn Schmitt.
Looking forward to great collaborative research!

Corrected figure...

10/10

Surprise bonus 🎁

Without explicit physics priors, 3DGSim also learns lighting & shadows 💡🕶️ as part of “dynamics” — showing its ability to model complex, diverse scene factors 🌍

9/10

Generalization? 🤔

Trained only on single-object-ground collisions, 3DGSim still handles multi-object interactions 📷 — preserving each object’s structure 🧱

8/10

Editability! ✏️🧩

Because 3DGSim uses an explicit 3D Gaussian representation, we can modify objects or environments mid-simulation—great for counterfactuals, scenario exploration, and modular setups.

7/10

Results!

3DGSim accurately simulates cloth, elastic, and rigid dynamics, capturing realistic motions and interactions across diverse scenarios.

6/10

We introduce three challenging datasets capturing distinct physical phenomena—rigid, elastic, and cloth dynamics (anchored at corners requiring implicit constraint learning)—each spanning unique interactions and deformation characteristics.

5/10

We extend MVSplat to support per-particle latent features, enabling latent 3D reconstruction—straight from multi-view RGB!

4/10

Powered by TEM-PTV3, our dynamic transformer hierarchically aggregates spatial information and systematically merges particle sets across timesteps, creating a unified, evolving 3D representation for simulations.

3/10
What’s novel?

3DGSim skips heavy biases (e.g., GNNs) & ground-truth 3D data (e.g., VPD). Training inverse rendering + dynamics end to end lets the encoder learn particle latents that capture both physical and visual aspects.

2/10
Website:
🔗 mikel-zhobro.github.io/3dgsim
#GaussianSplatting #Simulation #Inversegraphics #MachineLearning #ComputerVision

3DGSim encodes each frame as a 3D Gaussian particle set ✨, the dynamics transformer handles dynamics 🔄🌍, and 3D Gaussian splatting renders the scene 🎨

Ready to host GSO 2025 in Tübingen.
www.guided-self.org/gso-2025.html