Research

Robotic systems for manipulation, locomanipulation, and human–robot interaction

Grasp perception, motion planning, and operator interfaces for robotic and prosthetic control.

Language-guided grasping and execution on a legged manipulator.
Robot autonomy

Manipulation, locomanipulation, and robot learning

I study how mobile manipulators perceive and interact with objects when geometry is only partially observed. Current methods combine object segmentation and shape completion with 6-DoF grasp estimation, volumetric obstacle mapping, and collision-aware motion planning for reaching and grasping.

In our CoRL 2026 paper, we study the relationship between policy representations and locomotion smoothness during sim-to-real transfer. Related work conditions traversability estimates on the robot's morphology and physical capabilities.

Language-guided mobile manipulationTarget grounding, partial geometry, grasping Open-vocabulary shared autonomyGuidance, collision avoidance, valve tasks Capability-aware navigationEmbodiment-conditioned traversability Learning-based locomotionRepresentation health and sim-to-real control
Camera-based operator input and shared-control teleoperation of Spot's arm.
Human–robot interaction

Teleoperation, shared control, and intent decoding

I develop interfaces that translate operator motion and intent into robot commands. In camera-based teleoperation, RGB-D tracking estimates arm motion and hand gestures, while shared control assists target alignment, collision avoidance, and grasp execution. Gaze and mixed-reality interfaces provide additional ways to select targets and interact with robots.

For robotic and prosthetic hands, I use electromyography (EMG) and lightmyography to estimate gestures, hand motion, and grasping force. These studies examine discrete command recognition and continuous control from wearable measurements.

Vision-based robot teleoperationMotion mapping and shared grasp execution Wearable intent decodingEMG, lightmyography, gesture and force Gaze-oriented robot interactionMixed reality and assistive HRI
Systems and experiments

Selected research

Hardware implementations, experimental results, and system components.

Shared autonomy

Open-vocabulary shared autonomy for robotic manipulation

The operator selects a target with a free-form prompt and controls Spot through arm motion and hand gestures. The robot tracks the target, avoids collisions, and assists the final approach.

System components
  1. Calibration-free RGB-D teleoperation
  2. Qwen3-VL and SAM 2 target grounding
  3. nvblox volumetric collision mapping
  4. Potential-field assistance and cuRobo MPC
All shared-control trials succeeded18 cm minimum obstacle clearance
Shared autonomyOpen vocabularyMPC
Project page Paper Video
Mobile manipulation

Language-guided grasping under partial observation

The operator names an object in the scene. The system segments it, completes the partial RGB-D geometry, selects a collision-free 6-DoF grasp, and executes it on Spot.

System components
  1. Language request and open-vocabulary grounding
  2. Object-centric RGB-D reconstruction
  3. Execution-aware 6-DoF grasp selection
  4. Spot arm planning and execution
9/10 successful graspsReal-robot evaluationAccepted at SSRR 2026
VLM6-DoF graspingSpot
Publication page Paper Video
Field robotics

Capability-aware navigation across robot embodiments

The traversability model uses the robot's physical capabilities as an input. The same scene produces different maps for legged and wheeled platforms.

System components
  1. Scene-level visual representation
  2. Robot capability conditioning
  3. Platform-specific traversability prediction
  4. Real-time local obstacle avoidance
Accepted at IROS 2026Legged and wheeled robots
TraversabilityEmbodimentNavigation
Publication page Paper Video
Manipulation benchmarks

HiveBoard: Modular benchmarks for robotic and prosthetic manipulation

A modular, 3D-printed benchmark board for evaluating robot grippers, hands, and prostheses on industrial mechanisms, with reusable simulation assets.

System components
  1. Modular hexagonal benchmark board
  2. Torque, precision, and composed-assembly tasks
  3. 3D-printable hardware and articulated URDF/USD models
  4. Cross-platform evaluation protocol
13 interchangeable attachmentsFour manipulation platforms
Benchmark boardOpen hardwareManipulation
Project page Code
Human-guided control

Vision-based teleoperation of a legged manipulator

An RGB-D camera maps the operator's wrist motion to Spot's arm. Collision-aware trajectory planning assists arm control, and hand gestures trigger grasping.

System components
  1. Camera-based wrist pose estimation
  2. Real-time end-effector command mapping
  3. Collision-aware trajectory planning
  4. Gesture-triggered autonomous grasping
Best Paper · LARS 2025Real-time control
TeleoperationShared controlPose estimation
Publication page Paper Code Video

Related demonstration: lightmyography-based multigrasp robot-hand control. Source publication

Wearable sensing

Muscle–machine interfaces for intuitive robot control

EMG and lightmyography armbands decode gestures, in-hand motion, and grasping force for robot and prosthesis control.

System components
  1. Wearable EMG and optical sensing
  2. Gesture and force-intent decoding
  3. Real-time human-in-the-loop inference
  4. Robotic and prosthetic control
Scientific ReportsRA-L and TNSRE
EMGLightmyographyTransformers
Publication page Paper
Evaluation

Common experimental constraints

  • Real-time operation
  • Incomplete and noisy sensing
  • Safety and collision constraints
  • Evaluation on physical systems
Research environment

Platforms at the USP Center for Robotics

I work in the group led by Assoc. Prof. Marcelo Becker at the USP Center for Robotics.

Manipulation
KUKA iisy, Franka Research 3, DynaArm, NB 55, and Spot Arm
Legged and mobile
ANYmal-D, Spot, Unitree B2, Ascento, Taurob, TerraSentia, and custom platforms
Aerial
Fotokite, Voliro, Flybotix ASIO, and experimental UAV platforms