
Autonomous Robotic Manipulation Lab Robot Learning - Manipulation - Motion Planning #robotics Director: Dmitry Berenson https://nitter.cf/t.co/B1EAfHnUDZ
Ann Arbor, MI
Joined May 2022
- Tweets74
- Following57
- Followers251
- Likes62
We are excited to see you next week at #CoRL2026! We are presenting one conference paper and four workshop papers. Details below!
Workshop on Agentic Robotics
Topology-Informed Visual Prompting for Hierarchical Vision-Language-Action Control
Authors: @HaoyangWuX, @AbhinavKumarRob, Dmitry Berenson
1
1
23
Workshop: Modeling Uncertainty in Robotic World Models
Unified Belief-Conditioned World Modeling and Parameter Estimation for Multi-Finger Manipulation
Authors: @AbhinavKumarRob, Soshi Iba, Rana Soltani Zarrin, Dmitry Berenson
1
19
Excited to share MDF++: Visuo-Tactile World Action Model with End-to-end Representation Learning, led by recent ARM Lab graduate @ZixuanHuang15!
We'll be presenting at #CoRL 2026. Check out the thread below for more information.
1
3
140
MDF++ learns visuo-tactile representations end-to-end, outperforming existing World Action Models by up to 50% on contact-rich tasks. While naive joint training can lead to feature collapse, we introduce a multi-task learning strategy that leverages cross-modal correlations.
1
45
For more details, experimental results, and videos, check out our project website:
jointmdf.github.io/
If you're attending CoRL, come find us at Poster #158 on Tuesday, November 10, during Poster Session 2 (5:00–6:30 PM)!
26
We locally calibrate uncertainty in approximate dynamics models! By conditioning on latent observation embeddings, states, and actions, we enable uncertainty-aware planning in u̲n̲s̲e̲e̲n̲ test environments!
More on this @wafr_conf paper w/ @TheLuisMarques @UMRobotics below ⬇
Most conformal methods are input-independent, producing overconservative or overconfident plans. Our uncertainty-aware framework allows the planner to slow down when seeing ice while remaining fast in good conditions!
See: um-arm-lab.github.io/ocular/
4
76
Introducing STEMbot, coming soon to IROS 2026 in Pittsburgh!
Our miniature climbing robot combines compliant locomotion, geometric SLAM, semantic mapping, and manifold-constrained planning to autonomously navigate plant stems and branches.
Check out the project website for more information: um-arm-lab.github.io/stembot…
1
43
ARM Lab at University of Michigan retweeted
A 67-gram robot that climbs plant stems, builds a 3D map, and navigates autonomously among branches and leaves from @umicharmlab.
The goal: spot pests hiding under foliage before crop damage.
ICRA 2026 recap from the ARM Lab 🤖
We are excited to highlight our work: DTOUR Recovery, Multimodal Diffusion Forcing, Lies We Can Trust, and OCULAR.
Also check out work from lab alums Sheng Zhong and Mark Van der Merwe!
What an awesome week of robotics. Links below 👇
DTOUR Recovery @AbhinavKumarRob
dtourrecovery.github.io/
Multimodal Diffusion Forcing @ZixuanHuang15
unified-df.github.io/
Lies We Can Trust @TheLuisMarques
um-arm-lab.github.io/claps/
OCULAR @TheLuisMarques
um-arm-lab.github.io/ocular/
1
3
108
Estimating Deformable-Rigid Contact Interactions ieeexplore.ieee.org/document…
RUMI arxiv.org/abs/2408.10450
2
39
ARM Lab at University of Michigan retweeted
Given a dataset of multimodal robot trajectories, what is the most effective way to utilize the data?
In our new ICRA paper Multimodal Diffusion Forcing, we propose a unified multimodal diffusion model that is simultaneously a policy, world action model, dynamics model, and anomaly detector.
We demonstrate the capability of MDF on real world forceful manipulation tasks like engine maintenance.
arxiv: arxiv.org/pdf/2511.04812
Website: unified-df.github.io/
🧵
We propose a data-driven algorithm constructing sets guaranteed to contain future robot configurations at a user-defined probability, despite aleatoric and epistemic uncertainty. More details on this RAL publication below! w/ @TheLuisMarques @GhaffariMaani @UMRobotics
While current conformal prediction methods treat robot configurations as Euclidean, we analyze configuration errors in SE(2), leading to more volume-efficient sets.
Project: um-arm-lab.github.io/claps/
3
252





