Research

I study dexterous grasping and contact-rich manipulation under uncertainty using optimization, belief-space planning, and robot learning.

Research Overview

I study contact-rich poly-articulated dexterous grasping and manipulation under uncertainty. My research combines optimization-based control, kinodynamic motion planning, belief estimation, and policy learning to address uncertain object pose, friction, contact modes, and target geometry. A central question is how a robot can acquire and retain a grasp, establish and change contact, and complete manipulation while satisfying collision-avoidance and task constraints. I investigate these questions with fixed-base serial-chain manipulators, bimanual robot manipulators, and humanoids.

My earlier doctoral work addressed robust stochastic planning, safety-critical control using Boolean combinations of control barrier functions, distributional reinforcement learning, and distributed multi-agent coordination under uncertainty. These studies connect stochastic decision-making with safety and coordination constraints. The reinforcement learning studies include my risk-sensitive algorithms, reproductions of published methods, and integration of learned policies with robot feedback control.

current work

My grasping research addresses both grasp synthesis and feedback execution. I develop SE(3)-equivariant generative models for articulated hand configurations and object contacts, and differentiable belief representations for risk-sensitive grasp optimization under uncertain contact parameters. For grasp execution, I study feedback grasp selection, collision avoidance through control barrier functions, and preservation of grasp quality without a precomputed trajectory.

My manipulation research examines how uncertainty and changing contact conditions affect task completion. The model-based work includes risk-sensitive belief-space control for dexterous stowing and belief-informed hybrid control for bimanual assembly.

In ParcelStow, we compare expert and learned-policy success rates under variation in execution timing. My learning research includes ACT, diffusion policies, and residual reinforcement learning for bimanual manipulation. I also fine-tune vision-language-action models for monitoring task progress and triggering recovery from visuo-tactile feedback. A longer-term objective is to combine geometric structure, uncertainty estimation, and multimodal observations in manipulation policies that account explicitly for performance, robustness, and safety under partial observability.

I evaluate these methods in MuJoCo, Drake, and Isaac Sim, and through hardware experiments on the platforms listed on the Robots page.

Applications

Select a theme to browse its demonstrations; select a video to enlarge it.

Risk-Aware Dexterous Manipulation Grasp robustness, belief-space planning, and belief-informed hybrid control. 3 Demos

Robust Grasping & Dexterous Manipulation

Robust visuo-tactile dexterous grasping under uncertain object pose, friction, and contact modes.

Risk-Aware Manipulation

Risk-sensitive belief-space control for vision-guided dexterous stowing.

Belief-Informed Hybrid Control

Hybrid control law balancing information gathering and progress toward task-satisfying states, with almost-sure set-convergence guarantees.

Grasp Planning and Contact-Rich Manipulation Grasp synthesis, reactive grasp execution, visuo-tactile-guided planning, and contact-rich bimanual manipulation. 4 Demos

Whole-Body Dexterous Grasp and Manipulation Planning

DexReach is a GPU-accelerated Python package built on NVIDIA cuRobo for whole-body dexterous grasp and manipulation planning. The videos demonstrate grasping, pick-and-place, object reorientation, and bimanual manipulation in Isaac Lab.

Reach-Aware SE(3)-Equivariant Grasp Synthesis

Reach-aware, scene-aware, collision-free, and executable multifingered grasp generation.

Grasp Distance Fields

GDF supports autonomous dexterous grasp execution with dynamic obstacle avoidance and grasp-quality preservation.

Visuo-Tactile Bimanual Planning and Wrench-Aware Control

Autonomous bimanual assembly using depth-based pose estimation, collision-constrained joint-space trajectory optimization, tactile-informed and gravity-compensated position control, and wrench-aware control.

Robot Learning and Teleoperation Reinforcement and imitation learning, vision-language-action models, manipulation policy evaluation, and hardware teleoperation for contact-rich tasks. 8 Demos

Reinforcement Learning

I reproduced DAPG relocation results in MuJoCo / Adroit: 95/100 successful episodes with deterministic action selection and 98/100 with stochastic actions, compared with 0/100 for behavior cloning in both modes.

Imitation Learning

We compare expert and learned-policy success rates under variation in execution timing.

Foundation Policies · OpenVLA-7B

Supervised fine-tuning (SFT) of VLAs for monitoring task progress and triggering recovery from visuo-tactile feedback.

Safe Teleoperation

VR-assisted hardware teleoperation with single-arm workspace-boundary enforcement and two bimanual tasks, constrained object relocation and key–housing insertion. The insertion panel shows the left hand stabilizing the housing while the right hand seats the key.

Publications

2026

arXiv
Clinton Enwerem, Saleh Kemal, John S. Baras, and Calin Belta, “Improving Imitation Policies for Multi-Phase Tasks through Residual Learning Allocation,” arXiv preprint, 2026. arXiv link
arXiV
Clinton Enwerem, Saleh Kemal, John S. Baras, and Calin Belta, “Belief-Informed Hybrid Control with Almost-Sure Target-Set Convergence,” arXiv preprint, 2026. arXiv link
arXiV
Clinton Enwerem, John S. Baras, and Calin Belta, “Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds,” arXiv preprint, 2026. arXiv link
arXiV
Clinton Enwerem, John S. Baras, and Calin Belta, “Grasp Execution Without a Planner: Configuration-Space Grasp Distance Fields with Certified Safety & Guaranteed Quality,” arXiv preprint, 2026. arXiv link
arXiV
Clinton Enwerem, John S. Baras, and Calin Belta, “FIRMGrasp: A Friction-Informed Risk Margin for Robust Grasp Synthesis,” arXiv preprint, 2026. arXiv link
arXiV
Clinton Enwerem, John S. Baras, and Calin Belta, “EquiDexFlow: Contact-Grounded SE(3)-Equivariant Dexterous Grasp Generative Flows,” arXiv preprint, 2026. arXiv link
Peer-Reviewed
2026

CDC
Clinton Enwerem, John S. Baras, and Calin Belta, “Risk-Constrained Belief-Space Optimization for Safe Control under Latent Uncertainty,” To appear in the Proceedings of the 65th IEEE Conference on Decision and Control (CDC), 2026.
IROS
Clinton Enwerem, Shreya Kalyanaraman, John S. Baras, and Calin Belta, “Variational Neural Belief Parameterizations for Robust Dexterous Grasping under Multimodal Uncertainty,” In the Proceedings of the 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 3117–3124, 2026.
2025

CDC
Clinton Enwerem, Aniruddh G. Puranic, John S. Baras, and Calin Belta, “Safety-Aware Reinforcement Learning for Control via Risk-Sensitive Action-Value Iteration and Quantile Regression,” In the Proceedings of the 2025 IEEE 64th Conference on Decision and Control (CDC), pp. 4890–4895, 2025.
2024

CDC
Clinton Enwerem, Erfaun Noorani, John S. Baras, and Brian M. Sadler, “Robust Stochastic Shortest-Path Planning via Risk-Sensitive Incremental Sampling,” In the Proceedings of the 2024 IEEE 63rd Conference on Decision and Control (CDC), pp. 1087–1094, 2024.
ECC
Clinton Enwerem and John S. Baras, “Safe Collective Control under Noisy Inputs and Competing Constraints via Non-Smooth Barrier Functions,” In the Proceedings of the 2024 European Control Conference (ECC), pp. 3762–3768, 2024.
LCSS
Clinton Enwerem and John S. Baras, “Formation Tracking for a Class of Uncertain Multiagent Systems: A Distributed Kalman Filtering Approach,” IEEE Control Systems Letters, Volume 8, 2024.
2023

CoDIT
Clinton Enwerem and John S. Baras, “Consensus-Based Leader-Follower Formation Tracking for Control-Affine Nonlinear Multiagent Systems,” In the Proceedings of the 9th International Conference on Control, Decision and Information Technologies (CoDIT), 2023.

Earlier work (undergrad & early Ph.D.) →