Pratap Tokekar

Associate Professor at @umdcs, @umiacs, and @amazon Scholar. Previously @vt_ece @grasplab @umncomputersci

[Pratap Tokekar @ptokekar]

Multi-Agent Trust Region Policy Optimisation: A Joint Constraint Approach

Chak Lam Shek, @guangyao_shi, and Pratap Tokekar

Paper: https://t.co/BOcGB1gzLR #AAMAS2026

[Pratap Tokekar @ptokekar]

Two new papers at #CVPR2026 and #AAMAS2026 from @ruiliu0 and Chak Lam Shek from my lab. Rui and Chak are both on the job market this year!

[Pratap Tokekar @ptokekar]

Stable and Efficient Single-Rollout RL for Multimodal Reasoning
@ruiliu0 et al. #CVPR2026
Paper: https://t.co/emhB1Sz1Ll

[Rui Liu @ruiliu0]

I am heading to #CVPR2026 next week to present our work MSSR (Multimodal Stabilized Single-Rollout). This work aims to improve the training efficiency of multimodal RLVR by using just single rollout instead of group-based approaches such as GRPO.

With around 1/4 of training steps, our approach can reach similar accuracy as GRPO. While trained for the same amount of compute and test on multiple benchmarks for generalization performance, our approach can outperform other group-based baselines.

Thanks to all my amazing collaborators Dian Yu, Lei Ke, @HaolinLiu616, @YujunZhou0017, @LiangZhenwen, @haitaominlp, @ptokekar, Dong Yu.

I will also be actively looking for full time opportunities, very happy to connect!

Paper: https://t.co/h8FhLUrRQY
Project page: https://t.co/qOb2bYmV4r

[Pratap Tokekar @ptokekar]

Congratulations Dr. @peihong_yu Proud to be hooding and celebrating your well deserved success!

[JHU Computer Science @JHUCompSci]

Join us and @JohnsHopkinsIAA on Monday for a joint talk on trends in safe and reliable reinforcement learning, featuring @JHUAPL’s @aekoppel and @umdcs & @amazon Robotics’ @ptokekar. Learn more here: https://t.co/jAUsBHZrnt

[jodavaho @jodavaho]

Image with @ptokekar from @UMNCSE and @umdcs that I am obligated to repost occasionally.

[Amrit Singh Bedi @amritsinghbedi3]

Interested in theoretical RL?
Find us at poster #3312 at #NeurIPS2025

🗓 Thu Dec 4, 4:30–7:30 PM
📍 Exhibit Hall C/D/E

@anas__barakat @SOURADIPCHAKR18 @ptokekar @peihong_yu

[Pratap Tokekar @ptokekar]

@ruiliu0 @gammaumd Would especially love to chat with students interested in doing a PhD @umdcs and super cool work that we're doing at Amazon Robotics

[Pratap Tokekar @ptokekar]

Excited to be at #NeurIPS2025 this week. Would love to meet folks interested in robotics. We will be presenting two papers...

[Pratap Tokekar @ptokekar]

CAML: Collaborative Auxiliary Modality Learning for Multi-Agent Systems

Thu 4 Dec 4:30 p.m. PST — 7:30 p.m. https://t.co/GlczecfHLI
@ruiliu0 @gammaumd

[Pratap Tokekar @ptokekar]

‘Tis the season.. @jbhuang0604 has amazing advise as usual with really cool slides (including a blurb from me that’s so old it mentions something called Webex?!)

Some of my earlier notes in case you’re interested https://t.co/1SfXUFnRFU

[Jia-Bin Huang @jbhuang0604]

Sharing the slides for a talk on faculty job search

Hope it's helpful to people exploring and preparing for the process.

Feedback is welcome!

https://t.co/d0yWKt7sEO

[Pratap Tokekar @ptokekar]

We expanded our best paper finalist @DARSSymposium with a graph neural network based policy to restore connectivity in a team of robots facing adversarial conditions. Check it out!

[Autonomous Robots @AUROblog]

🆕Fast #kconnectivity restoration in #multirobotsystems for robust communication maintenance: algorithmic and learning-based solutions by @guangyao_shi, Md Ishat-E-Rabban, Griffin Bonner & @ptokekar.
🔗https://t.co/O5VEy1LhDx
@DARSSymposium @RAASumd @umdcs @iribecenter

[Pratap Tokekar @ptokekar]

Congratulations to Amisha on being a finalist in the student presentation competition at the Amazon Robotics research symposium!

[AmishaBhaskar @amishabhaskar]

Honored to be featured by @umiacs_umd for showcasing my work in robot learning at the @Amazon Robotics Research Symposium in Boston!

Grateful for my advisor @ptokekar, collaborators & the UMD community for their support.
🔗 https://t.co/7fyOyuQCt4

[Pratap Tokekar @ptokekar]

@mangahomanga @JohnsHopkins @JHUCompSci Welcome to the area!

[Pratap Tokekar @ptokekar]

Exciting new work on learning compact policies

[Anukriti Singh @singhanukriti]

Can robots get to the point?🎯
We’re sharing AFFORD2ACT—a way for robots to use semantics from a short prompt to keep only the pixels that matter and then act.
This replaces heavy, dense inputs with a small state for control.
1/6
#Robotics #AI #Manipulation #Affordance #Keypoints

[Pratap Tokekar @ptokekar]

Coming soon at #iccv25

[Kelin Yu @ColinYu14116982]

How can we utilize cross-embodiment action-free videos in learning generalizable and robust policies?

Introducing our paper in ICCV 2025 @ICCVConference (also best paper nomination at the ICRA FMNS WS), GenFlowRL: an reward shaping method with generative object-centric flow. 1/6

[Pratap Tokekar @ptokekar]

Coming soon at #NeurIPS2025

Cool work from @ruiliu0 on knowledge distillation for multi-modal multi-agent learning

[GAMMA UMD @gammaumd]

(1/n)

CAML: Collaborative Auxiliary Modality Learning for Multi-Agent Systems

Authors: @ruiliu0, Yu Shen, @penggao41252277, @ptokekar, @MingCLinCS
Link: https://t.co/iB98MdWF9Y

This Paper proposes Collaborative Auxiliary Modality Learning (CAML), a multi-modal multi-agent framework that enables collaborative training with shared modalities while supporting reduced-modality inference, achieving significant gains in accident detection for connected autonomous vehicles and semantic segmentation in aerial-ground robot data.

[Pratap Tokekar @ptokekar]

Coming soon at #NeurIPS2025

[Amrit Singh Bedi @amritsinghbedi3]

🚀 It feels refreshing to share some new theoretical results! #NeurIPS2025 🎉

🔑 One-line takeaway: We prove that policy gradient methods can achieve global optimality even in the broader setting of general utility RL

More details coming soon!!