ICRA'24: New results in SLAM and Control

Tactile Robot Programming

Authors [1] propose a framework for contact-rich manipulation control which allows to design controllers with minimal manual parameter tuning. The key observation is that the object-centric force and motion profile of a task dictates the task parameters in robot domain.

Robot Food Slicing

Authors [3] employ a "Forward Dynamics Compliance Controller" to execute RL-learned vegetable cutting policy. Notably, Controller's PD parameters are also inferred by an RL agent, not tuned manually. The research draws from the framework used in "Learning force control for contact-rich manipulation tasks with rigid position-controlled robots".

GP for Collision Checking

Authors [6] address the problem of autonomous agent's collision checking. Results are compared in terms of accuracy and speed against a strong baseline. What is notable is the accuracy of deriving the estimation variance.

New state estimation metric based on SE2(3)

Authors [7] propose a better single-value scoring system for estimated trajectories of poses. The benefits of above are: 1) absence of the need to balance rotatitional loss against translational loss 2) it better describes the discrepancy in velocity 3) allows for trajectory compression via spectral parametrization. This is losely based on previous work with preintegrated IMUs.

Enabling Passivity for Cartesian Workspace Restrictions

The article [8] follows the trend of energy-aware robotics, which ensure passivity of complex controller programs with use of the energy tank concept. This is applied to human-robot collaborative disassembly tasks. Several other passivity-based controllers are cited. NB: Kinematically redundant robots have more complex structure nullspace in the task Jacobian (it isn't static). Authors use `Bond Graph` figures from "Modeling and Control of Complex Physical Systems".

High-Precision Manipulator Position Tracking with ILC

Authors [11] solve for submillimeter precision of highly dynamic manipulator trajectories. They also cite other approaches to "shopfloor-ready" high-precision control schemes. An iterative learning controller (ILC) is used to mitigate unmodeled error terms (due to transmission dynamics errors, friction and elasticity in joints). Contains a reference to stability analysis of computed-torque controllers. Contains an illustrative diagram of inner and outer control loops.

Surface tracking with mobile base manipulator

Authors [12] provide many valuable references regarding hybrid control schemes, e.g. "An admittance-controlled wheeled mobile manipulator for mobility assistance: Human–robot interaction estimation and redundancy resolution for enhanced force exertion ability", "Adjusting the parameters of the mechanical impedance for velocity, impact and force control". They also explain in simple terms the control scheme implemented and the essence of nullspace for secondary tasks trick. The robot has a Wrist 6D Force/Torque Sensor.

NMPC with feedback

Authors [13] propose an NMPC that has probabilistic guarantees of feasiblity. The sampled trajectories are refined with feedback, which is computed via LQR. They explore a connection between policy-sampling MPC methods and RL control methods.

A scrubbing robots that cleans adhesive contaminants

In [15] authors avoided using ADRC (Active Disturbance Rejection Control) by implementing passive distirbance rejection with a tendon-driven robot. The deformation of the rubbing sponge is accounted for in accordance with Hooke's Law.

Inertia-based control for hitting known objects with manipulator

Authors [16] propose a control scheme that shapes manipulator's inertia to achieve consistent dynamic manipulation in varying configurations. Collision Mechanics is outlined. "Passive interaction control with dynamical systems" is cited. Authors describe a dynamics system which generates the pre-impact reference trajectory of the end effector. Authors build on their previous work "Learning to hit: A statistical dynamical system based approach". Robot's task space inertia matrix is computed from the robot's joint space mass matrix (Introduced by Khatib in "Inertial properties in robotic manipulation: An object-level framework"). "Manipulability of robotic mechanisms" is referenced in relation to "Manipulability Maximization Controller", which drives control towards a trajectory, that maximizes EE velocity in the goal direction. Anyone interested in nullspace projections for secondary tasks should check out "A consistent null-space based approach to inverse kinematics of redundant robots".

Reactive Trajectory Optimization for intercepting a thrown object mid-air

Authors [19] build upon the DPP (Differential Dynamic Programming) paradigm. "Probabilistic differential dynamic programming" is cited. Another good citation: "The Principle of Optimality in Dynamic Programming". Proposed method differs from DPP approaches in that the target state isn't a fixed value; it is probabilistic instead. Thus such a control sequence could be modeled probabilistically. The prior art on this includes "Gaussin Mixture Models for time series modeling, forecasting and interpretation" (2013), "Bayesian time series analysis" (2010). The key contribution of this article lies in solving for control corrections (formulated in Bayesian terms) to the initially computed DPP-trajectory. A result, unexpected by authors, is reported: longer optimization horizons lead to smaller control corrections. This is related to earlier work of [16], who also contributed to robotic projectile catching.

Filter-based SLAM with square-root update procedure

Authors [20] explore a novel SLAM architecture, not evaluated in earlier works. Prior to that, square-root update was demonstrated with Information-based Filter (e.g. iSAM2). Other covariance-based filters (e.g. MSCKF, MSCEqF) didn't use square-root update procedure. Publication is based on their previous result in SLAM enhancements, FEJ strategy. This result explored different existing marginalization schemes. Authors emhasize the importance of marginal covariance recovery in SLAM. OpenVINS is used to conduct a test in simulation.

Controller Tuning via Bayesian Optimization

Risk-averse & safe BO is explored in the article [22] in application to black-box tuning of a controller. Authors build upon their earlier work "Performance-based trajectory optimization for path following control using bayesian optimization". A monograph "Gaussian Processes for Machine Learning" is cited. Evaluated in simulation and with embodied 2d positioning system, tuning 4 variables: PID gains and the feedforward term.

Hybrid, Compliant Force-Impedance Control with fast motions

Proposed [23] system effectively tracks flat and curved surfaces with the set contact force. Authors mention indirect force control methods. They argue that with indirect force control the precise geometry model of the environment is required to ensure low contact forces. However, this precise geometry information is often unattainable. Another competing approach to control mentioned is "inverse dynamics control". Authors focus on tasks that have a force control component, which simulaneously being executed with high speeds. Thus the dynamics terms become non-negligible.

IKLink: EE motion planning of a serial maniputor

Authors [24] use an idea to employ dynamic programming on a graph to "stitch together" continuous segments of the target EE trajectory.

Learning from Demonstration with GMM and CBF

In [25], the safety aspect of a robot executing the teleoperated trajectory is explored. Authors state, that Dynamical Systems is a sound approach to LfD, but it's overly focused on convergence. The evolution of Dynamical Systems is bounded by Conic Control Barrier Function. The foundation of this work is the article "A physically-consistent bayesian non-parametic muxture model for dynamical system learning". Authors also published in 2023 a survey on DMPs (Dynamic Motion Primitives). Future work includes coupled Dynamical Systems.

Stereo-NEC

The work [28] deals with robust SLAM initialization in a IMU-aided, stereo setup. Earlier solutions exhibit an assumption of a) an adequate baseline between keyframes, selected for the init producedure, and b) moderate rotation. The proposed method, on contrary, is robust to pure-rotation initialization sequences. Important earlier work inludes MNEC (Finding the exact rotation between two images independently of the translation).

NeRF-VINS

Authors [30] aim to improve map-based localisation in absence of global information (loop closures, GPS). Loop Closure constraints are to be extracted from NeRF-generated novel views, corresponding to the current pose estimate. NeRF Hashing technique is used to cheaply generate the novel views on an edge device. Competing relocalisation solutions include an OSS maplab 2.0. Authors use MSCKF as a localisation framework, and besides the traditional state propagation with IMU data, and the measurement update with real images, another measurement update is introduced with a NeRF-generated stereo image, offset by 10 cm baseline from the current pose estimate.

Force-Feedback MPC

Contrary to many competitors, in [32] force-control is added to MPC framework without explicitely modeling the dynamics of contact forces. Authors cite "Real-Time motion planning of legged robots: A model predictive control approach", a successful recent MPC framework. Another notable reference: "Determining states of inevitable collision using reachability analysis".

Multi-Contact Feedback MPC

Listed under [33], cites Tedrake et al. on "Localizing external contact using proprioceptive sensors: The contact particle filter". The differentiating point of this work is the unplanned contact location identification.

Constrained Passive Interaction Control

Authors [34] propose a control achitecture which maintains passivity when a robot is in a feasible state; outside of it, it avoids self-collisions, any external collisions or singularities, and satisfies robot's kinematic limits. When a robot interacts with unknown environment, Lyapunov-stability is hard to apply, the passivity framework is used instead. One way to implement passivity with variable impedance control is a velocity field. The controller solves two tasks: implements tracking & passivity, and complies with 4 constraints, listed above. Control Barrier Functions are used in a QP controller.

Hand-eye calibration for a serial manipulator

Authors propose a robot-camera SE3 extrinsics calibration paradigm [36], where active perception is employed. End effector tracks a predetermined trajectory, consisting of multiple sphere arcs, rotated relative to one another. The camera tracks the "apogee" of each arc, and uses these points to solve for SE3 extrinsics. The approach is tested with da Vinci surgical robot. The calibration takes 1 minute in time and 1 cm^3 of space. A. Edsinger's thesis "Robot manipulation in human environments" is cited.

Miscellaneous

Learning to Play Foosball

Isaac Gym Sim + Real embodyment
of a robotic foosball player [2]

Pegasus

Quadruped-Wheeled underactuated mechanism [4]

Burrowing of a snake-like robot

LSTM-based policy informed by robot's altitude [5]

Online Policy RL driven by Uncertainty

Authors decouple epistemic and aleatoric uncertainity in the application of soft actor-critic network to a problem of bin picking [9].

Award-winning (IROS23) Cloth Unfolding solution

with custom built IR sensor; Relies upon several movement primitives [10].

A 2d robotic shoulder

A novel design [14]., which is different from competition in having self-sensing capabilities and not having bulky encoders.

Ground To Satellite Image Registration

A NN-based matching [17] for the ground-plane projected camera images with satellite views.

Smaller-footprint Visual Transformer for Relocalisation

Authors [18] conduct great experiments with recent competitors, and capture the current level of DNN-based relocalisation.

Topological Primitives used in manipulator's path planning

Authors [21] explore pulling, untangling, and threading environments for their proposed planning program. "MoveIt" and MPPI (Model Predictive Path Integral [Control]) were used in experiments. The modeling was carried out in Mujoco, and in real-world setup.

Vertical Vibratory Transport of Grasped Parts Using Impacts

An exciting design of a gripper capable of picking up various [26] objects with use of their inertia and gravity.

High performing SLAM solution with novel IMU-preintegration

Authors [27] include into IMU state current velocity, and use it in the pre-integration phase, allowing for better estimates during accelerating and decelerating motions.

Jerk-constrained trajectory planning in torque coordinates

A linearization scheme [29] was proposed to linearize 3rd order contraints on industrial manipulator's torque commands, which results in a time-optimal smooth trajectory.

A ball-intercepting control for a serial manipulator

A fast-reacting (~1/3 sec.) control program is proposed that itercepts a ball thrown into manipulator with its end effector. A database of pre-optimized trajectores was used. The proposed methods dynamically solves for faster and safer trajectories, resulting in 10x lower joint torques.

Wielding a screwdriver with a robotic hand

Authors [35] propose a hybrid force-control scheme that is capable of operating a common screwdriver.

How does strain-gauge sensor actually look like?

What goes into it [37] is a high-quality solid aluminum body (its responses to the applied stress are well modelled), and the sensors which read off these stresses during the real operation.

References

Kubra Karacan, Robin Jeanne Kirschner, Hamid Sadeghian, Fan Wu, and Sami Haddadin, Tactile Robot Programming: Transferring Task Constraints into Constraint-Based Unified Force-Impedance Control
Janosch Moos, Cedric Derstroff, Niklas Schröder, Debora Clever: Learning to Play Foosball: System and Baselines
Cristian C. Beltran-Hernandez, Nicolas Erbetti, Masashi Hamaya: SliceIt! - a Dual Simulator Framework for Learning Robot Food Slicing
Yuzhen Pan, Rezwan Al Islam Khan, Chenyun Zhang, Zhang Anzheng, Huiliang Shang: Pegasus: A Novel Bio-Inspired Quadruped Robot with Underactuated Wheeled-Legged Mechanism
Sean Even, Holden Gordon, Hoeseok Yang, Yasemin Ozkan-Aydin: Machine Learning-Driven Burrowing with a Snake-Like Robot
Javier Muñoz Mendi, Peter Lehner, Luis Moreno, Alin Albu-Schäffer, Maximo Roa: CollisionGP: Gaussian Process-Based Collision Checking for Robot Motion Planning
Varun Agrawal, Frank Dellaert: A Group Theoretic Metric for Robot State Estimation Leveraging Chebyshev Interpolation
Sebastian Hjorth, Johannes Lachner, Arash Ajoudani, Dimitrios Chrysostomou: Enabling Passivity for Cartesian Workspace Restrictions
Yitian Shi, Philipp Schillinger, Miroslav Gabriel, Alexander Qualmann, Hanna Ziesche, Zohar Feldman, Ngo Anh Vien: Uncertainty-Driven Exploration Strategies for Online Grasp Learning
Remko Proesmans, Andreas Verleysen, Francis wyffels: UnfoldIR: Tactile Robotic Unfolding of Cloth
Michael Schwegel, Andreas Kugi: A Simple Computationally Efficient Path ILC for Industrial Robotic Manipulators
Carlos Suarez Zapico, Y. R. Petillot, Mustafa Suphi Erden: Semi-Autonomous Surface-Tracking Tasks Using Omnidirectional Mobile Manipulators
Adam Polevoy, Marin Kobilarov, Joseph Moore: Probably Approximately Correct Nonlinear Model Predictive Control (PAC-NMPC)
Clemens Claudio Christoph, Amirhossein Kazemipour, Michel Ryan Vogt, Yu Zhang, Robert Kevin Katzschmann: Self-Sensing Feedback Control of an Electrohydraulic Robotic Shoulder
Noah Harmatz, Alina Zahra, Amir Abdelmalak, Shivam Purohit, Trevor Shin, Aaron Mazzeo: Hybrid Force-Position Control of an Elastic Tendon-Driven Scrubbing Robot (TEDSR)
Harshit Khurana and Aude Billard: Motion Planning and Inertia Based Control for Impact Aware Manipulation
Yanhao Zhang, Yujiao Shi, Shan Wang, Ankit Vora, Akhil Perincherry, Yongbo Chen, Hongdong Li: Increasing SLAM Pose Accuracy by Ground-To-Satellite Image Registration
Zhendong Xiao, Changhao Chen, Yang Shan, Wu Wei: EffLoc: Lightweight Vision Transformer for Efficient 6-DOF Camera Relocalization
Apan Dastider, Hao Fang, Mingjie Lin: RETRO: Reactive Trajectory Optimization for Real-Time Robot Motion Planning in Dynamic Environments
Chuchu Chen, Yuxiang Peng, Guoquan Huang: Fast and Consistent Covariance Recovery for Sliding-Window Optimization-Based VINS
Peter Mitrano, Dmitry Berenson: The Grasp Loop Signature: A Topological Representation for Manipulation Planning with Ropes and Cables
Christopher König, Miks Ozols, Anastasiia Makarova, Efe Balta, Andreas Krause, Alisa Rupenyan: Safe Risk-Averse Bayesian Optimization for Controller Tuning
Maged Iskandar, Christian Ott, Alin Albu-Schäffer, Bruno Siciliano, Alexander Dietrich: Hybrid Force-Impedance Control for Fast End-Effector Motions
Yeping Wang, Carter Sifferman, Michael Gleicher: IKLink: End-Effector Trajectory Tracking with Minimal Reconfigurations
Zheng Shen, Matteo Saveriano, Fares Abu-Dakka, Sami Haddadin: Safe Execution of Learned Orientation Skills with Conic Control Barrier Functions
Connor Yako, Jerome Nowak, Shenli Yuan, Kenneth Salisbury: Vertical Vibratory Transport of Grasped Parts Using Impacts
Yifu Wang, Yonhon Ng, Inkyu Sa, Alvaro Joaquin Parra Bustos, Cristian Rodriguez, Tao Jun Lin, Hongdong Li: MAVIS: Multi-Camera Augmented Visual-Inertial SLAM Using SE2(3) Based Exact IMU Pre-Integration
Weihan Wang, Chieh Chou, Ganesh Sevagamoorthy, Kevin Chen, Zheng Chen, Ziyue Feng, Youjie Xia, Feiyang Cai, Yi Xu, Philippos Mordohai: Stereo-NEC: Enhancing Stereo Visual-Inertial SLAM Initialization with Normal Epipolar Constraints
Jee-Eun Lee, Andrew Bylard, Zhouwen Sun, Luis Sentis: On the Performance of Jerk-Constrained Time-Optimal Trajectory Planning for Industrial Manipulators
Saimouli Katragadda, Woosik Lee, Yuxiang Peng, Patrick Geneva, Chuchu Chen, Chao Guo, Mingyang Li, Guoquan Huang: NeRF-VINS: A Real-Time Neural Radiance Field Map-Based Visual-Inertial Navigation System
Ramkumar Natarajan, Hanlan Yang, Qintong Xie, Yash Oza, Manash Pratim Das, Fahad Islam, Muhammad Suhail Saleem, Howie Choset, Maxim Likhachev: Preprocessing-Based Kinodynamic Motion Planning Framework for Intercepting Projectiles Using a Robot Manipulator
Armand Jordana, Sebastien Kleff, Justin Carpentier, Nicolas Mansard, Ludovic Righetti: Force Feedback Model-Predictive Control Via Online Estimation
Seo Wook Han, Maged Iskandar, Jinoh Lee, Min Jun Kim: Online Multi-Contact Feedback Model Predictive Control for Interactive Robotic Tasks
Zhiquan Zhang, Tianyu Li, Nadia Figueroa: Constrained Passive Interaction Control: Leveraging Passivity and Safety for Robot Manipulators
Ling Tang, Yan-Bin Jia, Yuechuan Xue: Robotic Manipulation of Hand Tools: The Case of Screwdriving
Billy Zhong, Bin Li, Wei Chen, Yunhui Liu: Robot-Camera Calibration in Tightly Constrained Environment Using Interactive Perception
Takamasa Kawahara, Toshiaki Tsuji: Development of an Easy-To-Cut Six-Axis Force Sensor