We’re really excited about the speakers we have lined up, including Animesh Garg (Georgia Tech) @animesh-garg.bsky.social , Daniel Ho (1X Technologies), Hao Su (UCSD), Katerina Fragkiadki (CMU), Yilun Du (Harvard) @yilundu.bsky.social, and Russ Tedrake (TRI + MIT).

This was led by Albert Wilcox with support from Mohamed Ghanem, Masoud Moghani, Pierre Barroso, Benjamin Joffe and myself.
.
Learn more:
🌐 Website: pairlab.github.io/Adapt3R
📄 Paper: arxiv.org/abs/2503.04877
🖥️ Code: github.com/pairlab/Adap...

On our real-world multitask IL benchmark, Adapt3R achieves a strong in-distribution success rate and sees the smallest performance loss by far under a dramatically new viewpoint.

Here we rotate the scene camera by θ radians about the EE starting position, and see that Adapt3R consistently achieves the strongest performance. Notably, it maintains >80% success rate on LIBERO, and has the only nonzero MimicGen success rate when θ ≥ 0.6.

With an unseen embodiment, Adapt3R consistently outperforms its 2D counterparts and DP3, and outperforms 3D Diffuser Actor in two of our four settings. 2D features lifted into 3D are an effective representation for this scenario and Adapt3R makes good use of them!

When we evaluate policies trained with Adapt3R on the multitask LIBERO benchmark or the high-precision tasks from the MimicGen paper, we see that they are just as performant as their RGB counterparts.

Adapt3R unprojects 2D features into a point cloud, transforms them into the end effector’s coordinate frame, and uses attention pooling to condense them into a single conditioning vector for IL. Notice that Adapt3R attends to the same points before and after the camera change!

Learning 3D representations is hard without 3D data
💡 The key idea is to use a 2D foundation model to extract semantic features, and use 3D information to localize those features in a canonical 3D space without extracting any semantic information from the 3D data.

Albert Wilcox has been working on the using canonical 3D reps instead.
Yet, naive 3D alternatives don't work since most of the data is not easily featurized.

Adapt3R is a 3D backbone that works with your favorite robot learning method and generalizes to unseen embodiments & camera viewpoints!

@unitreerobotics.bsky.social great job.
Looking forward to more specs on the robot.

Are we there yet? No!
Humanoids will need all the human help they can get

I believe working on this problem might well be one of the coolest things you'll ever do!

I am leading our AI efforts for Humanoids.
I will elaborate in a blogpost in the upcoming weeks.

for now...to the stars and beyond!🚀

It's a multifaceted decadal challenge.
There is a palpable excitement with tremendous progress - we are moving faster than ever before.

Apptronik has the ambition & heft to take on one of most impactful problems of this decade!
Thrilled to join Jeff Cardenas & Nick Paine on this journey

Check out our website any-place.github.io & paper www.arxiv.org/abs/2502.04531 for more details!

📣 to all our authors & collaborators: @allanzhao.bsky.social, Miroslav Bogdanovic, Chengyuan Luo, Steven Tohme, Kourosh Darvish, @aspuru.bsky.social, Florian Shkurti, @animesh-garg.bsky.social [5/5]

🔍 Check out more at
any-place.github.io
www.arxiv.org/abs/2502.04531

Yuchi Zhao Miroslav Bogdanovic Chengyuan Luo Steven Tohm Kourosh Darvish @aspuru.bsky.social Florian Skhurti & Animesh Garg

@uoftcompsci.bsky.social
@vectorinstitute.ai @accelerationc.bsky.social @gtresearchnews.bsky.social

How well does AnyPlace perform?
🏆 Simulation results: Outperforms baselines in
✔ Success rate
✔ Coverage of placement modes
✔ Fine-placement precision
📌 Real-world results: Our method transfers directly from synthetic to real-world tasks, succeeding where others struggle!

To generalize across objects & placements, we generate a fully synthetic dataset with:
✅ Randomly generated objects in Blender
✅ Diverse placement configurations (stacking, insertion, hanging) in IsaacSim
This allows us to train our model without real-world data collection! 🚀