Joseph Amigo

PhD Candidate at NYU & LAAS-CNRS - Robotics, Reinforcement Learning, Deep Learning

[Chris Paxton @chris_j_paxton]

CoRL has become the place to be if you are interested in robotics. There's no robotics without intelligence -- robots will never scale without scalable learning for problems like contact-rich mobile manipulation -- and so we've seen this academic conference become so essential IMO. Looking forward to meeting people there

[Jerry Chéng @thejerrycheng]

Today, I'm excited to finally introduce MABEL — our fully open-source, mobile bimanual robot! 🤖🦾

$8K in parts. Just 2 weeks from parts to a fully tested robot.

And we're open-sourcing EVERYTHING: hardware, firmware, teleop pipeline, controllers, learning pipelines, and digital twin simulation.

Built by us at @nyuniversity and @UofTRobotics in 3 months, started in a tiny NYC studio apartment 🗽🍕

Our goal is simple: make capable, full-stack robotics accessible to everyone. 🚀
🧵👇
(1/7)

[Joseph Amigo @Jsphamigo]

@qiayuanliao @breadli428 Separately, I think the reward landscape matters: first-order gradients shine on smooth objectives like dense tracking, zeroth-order is more robust on non-smooth/sparse ones, and contacts are where first-order gets delicate.

[Joseph Amigo @Jsphamigo]

@qiayuanliao @breadli428 Agree the gap depends a lot on implementation choices, so controlled comparisons are hard.

[Joseph Amigo @Jsphamigo]

@qiayuanliao @breadli428 Building on Qiayuan's point: since DMO keeps every rollout in the simulator and only uses the network for the backward pass, it's a natural controlled setting for this question.

[Qiayuan Liao @qiayuanliao]

I think your claim conflicts with both SuperMimic and DMO, @Jsphamigo.
In SuperMimic, we have an ablation showing that APG is better with the common standard whole-body tracking setup.

DMO has a more interesting setup. It has the same env/task as RWM (quadrupedal locomotion without reference motions). But for sampling and gradients, it does the exact opposite of RWM: the forward rollout uses simulation, but the backward pass uses the first-order gradients from a learned network.

Indeed, we might need to run more controlled experiments.

[Joseph Amigo @Jsphamigo]

@qiayuanliao Thanks Qiayuan, really appreciate it! Congrats again on SuperMimic 🙌

[Qiayuan Liao @qiayuanliao]

Just realized that we didn’t discuss the DMO in our paper! The DMO provided lots of insight when we were working at SuperMimic. It tells us how powerful a first-order gradient is by construction, which can complete the only missing term in the 2x2 ablation grid. That is
{PPO, APG}x{sim, WM}.

[Joseph Amigo @Jsphamigo]

@qiayuanliao For context: DMO (CoRL '25) showed the same two effects on a single V100. First-order gradients through a learned model needed far fewer environment steps than zeroth-order RL (PPO/SAC) and converged faster in wall-clock time 👇

[Joseph Amigo @Jsphamigo]

Great to see first-order gradients through learned world models scale to whole-body tracking. Congrats @qiayuanliao et al.! 👏
We explored a related, decoupled variant in DMO (CoRL '25): forward in the real simulator, backward through a learned model. https://t.co/SrLCn1mdVA ⁠

[Qiayuan Liao @qiayuanliao]

We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.

Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.

Project: https://t.co/6gc9WKm9qY
Demo: https://t.co/OHRvx5ogQY

[Joseph Amigo @Jsphamigo]

@thejerrycheng Thank you so much! Fully agree, world models are a big part of the recipe for robotics. Great to see more people jumping in!

[Chenhao Li 📌 CoRL @breadli428]

It’s great to see growing interest in MBRL. Last year in RWM, we demonstrated that policies trained entirely in world-model imagination transfer zero-shot to the real world.

Sometimes, ideas are ahead of their time. Glad to see this direction getting the recognition it deserves.

[Dyuman Aditya @dyumanaditya]

🤖 Ever wondered if you can train end-to-end humanoid policies entirely within a differentiable simulator?

🥁 Introducing 𝗕𝘂𝗻𝗱𝗹𝗲𝗱 𝗖𝗼𝗻𝘁𝗮𝗰𝘁 𝗚𝗿𝗮𝗱𝗶𝗲𝗻𝘁𝘀 (𝗕𝗖𝗚)

🌐: https://t.co/h4i7NcBD0U
📄: https://t.co/b7TAjE8Yhh

🧵👇

[Unitree @UnitreeRobotics]

Unitree Breakthrough: The World’s First Real-Time World Model-Driven Fully Autonomous Humanoid Robot Combat🥊
UnifoLM-X2-1.0 breaks through world-action foundation models' bottlenecks in instant planning, decision-making, and dynamic interactive execution, achieve high dynamics, strong interaction, real-time prediction and planning of the future, achieve fully autonomous combat for humanoid robots. This validates the fundamental feasibility of large-scale deployment of world model-driven humanoid robots.

[Jerry Chéng @thejerrycheng]

Come checkout my lab mates great work on using world model for RL policy! They trained their interactive video model completely from scratch - the best I have ever tried.

https://t.co/EjgK7ZJPXV

[Frontier Signal @FrontierSignal_]

@chris_j_paxton @Jsphamigo @Rk4342R Clever split, but the local model still needs feeding. Contact dynamics come from real contact hours, so the bottleneck moves from simulator fidelity to collecting messy physical data.

[Don Burnette @don_burnette]

@chris_j_paxton @Jsphamigo @Rk4342R Super cool! Agree this is the tractable path. Contact dynamics are high frequency and short range, so paying for resolution only where the fingers touch is where the compute should go.

[RoboPapers @RoboPapers]

Reinforcement learning has unlocked tremendous gains in broader AI and humanoid robot whole body control, but still has yet to deliver strong results for robot manipulation due to the difficulty of capturing tasks in an accurate simulation; one way to fix this is to create complex environments with learned world models. But learning fully within world models is often intractable as powerful world models are computationally too expensive.

Instead, @Jsphamigo and @Rk4342R propose to break the problem down: into a large-scale world model that can generate forward trajectories, and a lightweight, low-dimensional latent-space model which can approximate local dynamics of a problem, without needing to back-propagate through a heavy global model. What this means: you can do reinforcement learning in this coupled local-global world model, and learn contact rich skills for a robot.

To learn more, watch Episode #101 of RoboPapers, with @micoolcho and @DJiafei!

[Chris Paxton @chris_j_paxton]

@Jsphamigo @Rk4342R Learn more

Project page: https://t.co/b3Jt15Uos0

ArXiV: https://t.co/hNLjvrrDvP

This on substack: https://t.co/fXNzpR6M74

[Chris Paxton @chris_j_paxton]

Learn contact-rich humanoid manipulation policies, purely within a world model. This is usually not tractable, because you need a very large model to get the contact dynamics right -- instead, @Jsphamigo @Rk4342R couple a large global model with a local model. Very cool work on @RoboPapers ->