iiinova — where your digital self keeps evolving

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

And thanks to all involved PIs: @jan-peters.bsky.social, @chelseafinn.bsky.social, Hao Su, Henrik I. Christensen, @georgiachal.bsky.social, Marc Schwager and Marco Hutter!

020

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

Thanks to my co-lead Bo Ai! Thanks to everyone already helping with compute and deployment: Dichen Li, Tongzhou Mu, Sophie Lueth, Nicolas Hahn, John Tucker and @breadli428.bsky.social !

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

👥 Meet the growing team, watch the real-world deployments and find out how to get involved on our website: gamma-zero.com

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

🤝 Accepted robot models are cited or acknowledged. Contributions to training and deployment on new robots may qualify for authorship. Send your model or get in touch about deployment by October 15: contribute@gamma-zero.com

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

🛠️ Help us grow the open robot collection! Have a URDF of your robot? With high-quality meshes, inertias, joint limits, and actuator torque and velocity limits? We'd love to include it! Have hardware in your lab? Help us train and deploy on robots we haven't tried yet.

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

🤖 Try it in your browser, explore different robots, change their physical properties, or upload your own robot! Website: gamma-zero.com Robot playground: play.gamma-zero.com

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

⚡ One policy, millions of embodiments, over 200 robot models. Can we add yours? We're building γ₀, a generalist RL policy for motion control trained across millions of randomized embodiments derived from a growing collection of more than 200 robot models.

192

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

🙌 And also lots of thanks to Siwei Ju that helped a lot on the deployment from the @ias-tudarmstadt.bsky.social side, while I was already in SF, and @jan-peters.bsky.social for supporting us as always!

020

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

🤝 I'm always super happy to collaborate with good friends and colleagues like An Thai Le! The Vietnam and Darmstadt connection stays strong!

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

🙏 Congrats to Tan-Dzung Do, Tuan Dat Phuong, Cuc Trinh, Vien Anh Ngo, and Xinchao Wang for pushing this.

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

🔀 Having worked a lot on multi-embodiment learning, I love how simple the idea is and we are finally getting closer and closer to learning fully transferable multi-embodiment motion tracking!

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

🤖 A BFM lets one latent vector prompt a humanoid to track a motion, reach a pose or maximize a reward. 🚀 CrossBFM transfers the latent space to any humanoid in just an hour.

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/10/2026

⚡️ Happy to share CrossBFM, our new work on cross-embodiment Behavior Foundation Models! 🔗 Project page: dotandung.github.io/crossbfm/

161

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2026

Checkout the baselines this work is inspired by: - SRL: arxiv.org/abs/2605.31273 - CRL: arxiv.org/abs/2503.14858 - AC-CRL: arxiv.org/abs/2608.30640

110

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2026

🚧 Limitations 2: Also the JaxGCRL policies are not optimal. Sometimes strange gaits and generalization to changing goals in episodes does not work well. Lots of room for future work!

110

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2026

🚧 Limitations 1: Self-supervised RL is still significantly worse than reward-based RL. PPO is far better at velocity-tracking, the Go2 climbs boxes backward, and on the Unitree G1 our policies did not learn to walk at all.

110

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2026

🐕 To test the limits of self-supervised RL, we built velocity tracking, goal-position reaching and box climbing tasks for the Unitree Go2 ChronoSRL is the only method that learns to stay at the commanded velocities and goal positions, and it climbs the highest boxes

110

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2026

🚀 On JaxGCRL tasks, ChronoSRL learns faster than Contrastive RL, Action Chunking + CRL and Survival RL across network depths ⚡️ With 8+ layers, it reaches the strongest baseline's final performance in 30M steps 🪶 In ant mazes, 4 layers already learn faster than any baseline with up to 64 layers

110

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2026

🔑 The key idea: Train the distance between state-action and goal embeddings to match the goal-reaching time, and push unreached goals at least one discount horizon away ⏱️ Building on survival RL, two heads then predict when the agent arrives and how long it stays near the goal

110

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2026

⚡️ A goal that is close in space can be far away in time ⏳ Introducing ChronoSRL: We give the critic of self-supervised RL a temporal geometry, so embedding distances measure the time to reach a goal 🤖 Play with the policy on quadruped task and more: nico-bohlinger.github.io/chronosrl_we...

122

Nico Bohlinger @nicobohlinger.bsky.social · 12/09/2026

Happy to share that I joined Amazon FAR (Frontier AI & Robotics) in San Francisco as a Research Intern! 🦾 Really looking forward to exciting research around humanoids and reinforcement learning, under supervision of @carlosferrazza.bsky.social and Rocky Duan.

061

Nico Bohlinger @nicobohlinger.bsky.social · 01/09/2026

Finally at the International Workshop of Intelligent Autonomous Learning Systems (IWIALS) 2026 in Austria organized by Erik Helmut, Michael Drolet, @jan-peters.bsky.social from @ias-tudarmstadt.bsky.social .

010

Nico Bohlinger @nicobohlinger.bsky.social · 01/09/2026

Then at the IJCAI 2026 Workshop on AI-Based Humanoid Robot Design and Control Through the Lens of HRI, Evolution, and Biomechanics organized by Melya Boukheddimi from @dfki.bsky.social Bremen

110

Nico Bohlinger @nicobohlinger.bsky.social · 01/09/2026

First at the Real-World RL Workshop in Hannover organized by Marius Lindauer from LUH-AI

110

Nico Bohlinger @nicobohlinger.bsky.social · 01/09/2026

⚡ In the last few weeks, I had the great chance to present our recent work on embodiment-aware reinforcement learning for robot control and design 🤖

141

Reposted by Nico Bohlinger

Kai Ploeger @kaiploeger.bsky.social · 24/06/2026

Learning five-ball juggling on the second attempt, with two Barrett WAMs. Most humans take years of practice. Paper and videos: kai-ploeger.com/residual-juggling #robotics #juggling

145

Nico Bohlinger @nicobohlinger.bsky.social · 16/06/2026

I had a great time visiting Prof. Kevin S. Luck at @vuamsterdam.bsky.social and to present our recent works on embodiment-aware learning for robot control and design! So much cool work here in Amsterdam about Robot Co-Design! There is huge potential in the data-driven way of designing robots!

021

Nico Bohlinger @nicobohlinger.bsky.social · 01/06/2026

P.S. I'm at ICRA in Vienna this week to present: Active Embodiment Identification with Reinforcement Learning for Legged Robots @ Workshop: Act to Sense to Act Better(arxiv.org/pdf/2605.08020) If you want to chat about anything related to multi-embodiment, co-design, locomotion, RL, etc. hit me up😄

arxiv.org

020

Nico Bohlinger @nicobohlinger.bsky.social · 01/06/2026

🔗 Check out the project page: nico-bohlinger.github.io/shape-your-b... 📄 Read the full paper: ias.informatik.tu-darmstadt.de/uploads/Team... 📹 Watch the video: youtu.be/_SHSWUSQZvg

nico-bohlinger.github.io

Shape Your Body

Shape Your Body: Value Gradients for Multi-Embodiment Robot Design

220

Nico Bohlinger @nicobohlinger.bsky.social · 01/06/2026

🛠️ These value gradients do more than just optimize: they can act as a design analysis tool. 🔍 The gradients could tell robot engineers exactly where to look first to fix bottlenecks in the design, whether that means adjusting body masses, joint axes, foot geometry or even control gains.

120

Nico Bohlinger @nicobohlinger.bsky.social · 01/06/2026

🤖 So, what do the optimized robots look like? Starting from perturbed designs, VGDS successfully reshapes the embodiment to improve the frozen policy's return. For tall humanoids, the gradients consistently reshape them to be smaller, wider, and more compact to improve stability!

130

Nico Bohlinger @nicobohlinger.bsky.social · 01/06/2026

⏱️ The biggest win is computational efficiency Adapted RL co-design methods scale linearly because they require a new training run per design. With VGDS, after one initial training run, each additional design takes only 1–2 minutes of search!

130

Nico Bohlinger @nicobohlinger.bsky.social · 01/06/2026

🚀 To make the value function a true generalist, we scaled up multi-embodiment training to 50 robots. Using the Unified Robot Morphology Architecture (URMA), the policy and critic learn across 15 quadrupeds, 31 bipeds/humanoids, and 4 hexapods. ⚙️ Each robot has 190 to 1177 design parameters.

130

Nico Bohlinger @nicobohlinger.bsky.social · 01/06/2026

🔑 The key idea: Train one embodiment-aware policy + value function across many robots, then reuse the frozen value function as a differentiable surrogate for robot design. Our search method, VGDS, follows value gradients while staying close to a reference design through a soft trust region.

130

Nico Bohlinger @nicobohlinger.bsky.social · 01/06/2026

⚡️What if we could design robots with gradients? 🤖 Introducing Shape Your Body: we train one multi-embodiment policy + value function, then optimize new robot designs through value gradients. 🔗 Try out our interactive demo here: nico-bohlinger.github.io/shape-your-b...

141

Nico Bohlinger @nicobohlinger.bsky.social · 20/10/2025

4. Multi-Embodiment Locomotion at Scale with extreme Embodiment Randomization -> Monday Workshop on Foundation Models for Robotic Design & Friday Workshop on Challenges and Applications Prospects for Paper: ias.informatik.tu-darmstadt.de/uploads/Team...

010

Nico Bohlinger @nicobohlinger.bsky.social · 20/10/2025

3. Evaluation of an Actuated Spine in Agile Quadruped Locomotion -> Friday Workshop on Climbing Robotics Paper: ias.informatik.tu-darmstadt.de/uploads/Team...

ias.informatik.tu-darmstadt.de

110

Nico Bohlinger @nicobohlinger.bsky.social · 20/10/2025

2. Bridge the Gap: Enhancing Quadruped Locomotion with Vertical Ground Perturbations -> Wednesday Session, Bioinspired Robot Learning Website: nico-bohlinger.github.io/bridge_the_g...

110

Nico Bohlinger @nicobohlinger.bsky.social · 20/10/2025

1. Gait in Eight: Efficient On-Robot Learning for Omnidirectional Quadruped Locomotion -> Wednesday Session, Legged Robots 1 - Locomotion Website: nico-bohlinger.github.io/gait_in_eigh...

110

Nico Bohlinger @nicobohlinger.bsky.social · 20/10/2025

I'm presenting four different works at IROS 2025 this week in Hangzhou 🤖

152

Nico Bohlinger @nicobohlinger.bsky.social · 02/10/2025

If you are interested in massive multi-embodiment learning, come and chat with me at: - Today, WS Sim-to-Real Transfer for Humanoid Robots at Humanoids2025 - Oct 20th, WS Foundation Models for Robotic Design at IROS2025 - Oct 24th, WS Reconfigurable Modular Robots at IROS2025

110

Nico Bohlinger @nicobohlinger.bsky.social · 02/10/2025

⚡️ Can one unified policy control 10 million different robots and zero-shot transfer to completely unseen robots, even humanoids? 🔗 Yes! Checkout our paper: arxiv.org/abs/2509.02815

120

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2025

👏 Huge thanks to everyone involved: Bo Ai, Liu Dai, Dichen Li, Tongzhou Mu, Zhanxin Wu, K. Fay, Henrik I. Christensen, @jan-peters.bsky.social and Hao Su

020

Nico Bohlinger @nicobohlinger.bsky.social · 30/09/2025

🇰🇷 Conferences are about finally meeting your collaborators from all around the world! Check out our work on Embodiment Scaling Laws @CoRL2025 We investigate cross-embodiment learning as the next axis of scaling for truly generalist policies 📈 🔗 All details: embodiment-scaling-laws.github.io

292

Reposted by Nico Bohlinger

Antonin Raffin @araffin.bsky.social · 07/07/2025

Need for Speed or: How I Learned to Stop Worrying About Sample Efficiency Part II of my blog series "Getting SAC to Work on a Massive Parallel Simulator" is out! I've included everything I tried that didn't work (and why Jax PPO was different from PyTorch PPO) araffin.github.io/post/tune-sa...

araffin.github.io

Getting SAC to Work on a Massive Parallel Simulator: Tuning for Speed (Part II) | Antonin Raffin | Homepage

This second post details how I tuned the Soft-Actor Critic (SAC) algorithm to learn as fast as PPO in the context of a massively parallel simulator (thousands of robots simulated in parallel).

4358

Nico Bohlinger @nicobohlinger.bsky.social · 02/07/2025

Or come to my talk @ International Symposium on Adaptive Motion of Animals and Machines and LokoAssist Symposium (AMAM) on Friday at TU Darmstadt Thanks to @ias-tudarmstadt.bsky.social, @jan-peters.bsky.social

051