Abhinav Agarwal

Research Intern at @Meta | Robotics @MIT | Former @princeton MAE

[Abhinav Agarwal @_abhinav_ag]

Roboticists have been working hard to add memory to “goldfish” imitation learning policies. But do we understand how context length impacts state-of-the-art imitation learning methods?

We provide an in-depth analysis of long-context Diffusion Policies to study this

🧵(1/7)

[Adam Wei @adamwei_]

🎉 Ambient Diffusion Policy was accepted to CoRL 2026! See y'all in Austin 🫡

[Abhinav Agarwal @_abhinav_ag]

@likeavass @RussTedrake Agreed that long execution horizons are compensating for expert non-Markovianity wrt policy input. Important: results specifically focus on open-loop “execution”. Chunked prediction could still have benefits (clarifying as MPC thinking focuses on planning horizon as well)

[Abhinav Agarwal @_abhinav_ag]

@roblee_rl @michzeng_ Great idea. It’ll require careful experimentation to try, especially because I have heard of concerns that the policies can overfit to proprioception history (and learn to ignore RGB conditioning). We haven’t tried it here; certainly curious about it

[Michael Zeng @michzeng_]

Action chunking — especially executing long action sequences open-loop — is widely used in imitation learning for robotic manipulation. Why is it so effective and do we really need it? We find a key reason:

Long open-loop execution helps short-context policies imitate non-Markovian experts.

With this insight, we show how to move beyond open-loop execution: extending policy context restores reactivity while achieving even higher task performance.

🧵(1/5)

[Abhinav Agarwal @_abhinav_ag]

Behavior cloning’s success has been accompanied by design choices we don’t understand well. Action chunking with open-loop execution is one of them.

We provide a key reason why open-loop execution seems important: it helps short-context policies imitate non-Markovian experts.

Important practical takeaway: focusing on increasing relatively recent context length can help restore reactivity.

Congratulations @michzeng_ for leading this insightful work.

[Michael Zeng @michzeng_]

Action chunking — especially executing long action sequences open-loop — is widely used in imitation learning for robotic manipulation. Why is it so effective and do we really need it? We find a key reason:

Long open-loop execution helps short-context policies imitate non-Markovian experts.

With this insight, we show how to move beyond open-loop execution: extending policy context restores reactivity while achieving even higher task performance.

🧵(1/5)

[Abhinav Agarwal @_abhinav_ag]

Good point. In our paper we use “DiT��� as a shorthand for diffusion transformer baseline from the original DP work (w/ cross-attn conditioning). It has been used in other long context DP learning works: https://t.co/8jawRMTWcw. So using it allowed us to compare. Agreed that we need more policy architecture evaluations as a community.

[Abhinav Agarwal @_abhinav_ag]

Roboticists have been working hard to add memory to “goldfish” imitation learning policies. But do we understand how context length impacts state-of-the-art imitation learning methods?

We provide an in-depth analysis of long-context Diffusion Policies to study this

🧵(1/7)

[Abhinav Agarwal @_abhinav_ag]

Paper: https://t.co/GeaU3uL9yi
Website: https://t.co/KChxulZqW5

Thank you to my co-authors: @adamwei_ , @tkargin_ , @michzeng_ , @colemobecker , Kerem Dayi, Pablo Parrilo, @AsuOzdaglar , @RussTedrake

🧵(7/7)

[Abhinav Agarwal @_abhinav_ag]

We find that long-context policies can be especially sensitive to vision encoder training - comparing against methods from prior works, we find that freezing the vision encoder can help us succeed in many cases with "past-token prediction", but there are examples where this fails.

🧵(6/7)

[Adam Wei @adamwei_]

🤖 We introduce Ambient Diffusion Policy, a simple and principled method for training policies with suboptimal data in robotics.

Suboptimal data is everywhere in robotics…
❌ Data filtering is wasteful
❌ Co-training learns both good and bad features
✅ Ambient Diffusion Policy selectively learns useful features via noise-dependent data usage

👇🧵(1/5)

[Adam Wei @adamwei_]

Happy to announce that we’ll be presenting our work on sim-and-real cotraining at IROS 2025!

Check out our latest arXiv version - we’ve added new experiments for 2 alternative co-training formulations and addressed some FAQs

📄 https://t.co/ezpVyOOfpG
🌐 https://t.co/jW05A7uwBU

[Abhinav Agarwal @_abhinav_ag]

Exciting results that will help us understand and design effective co-training methods in robot imitation learning. Grateful to have worked with brilliant collaborators!

[Adam Wei @adamwei_]

Learning from both sim+real data could scale robot imitation learning. But what are the scaling laws & principles of sim+real cotraining?

We study this in the first focused analysis of sim+real cotraining spanning 250+ policies & 40k+ evals https://t.co/hMc56Hj2K9 (1/6)