How I didn't go to IROS'23

Just reading through, let alone summarizing and analyzing the proceedings of the whole conference is an effort beyond the resouces of a single person, so this review is focused on just a few topics to the author's preference:

  • NeRF
  • Manipulation
  • Miscellaneous

Some results receive no detailed comment here due to a lacking in the author's expertise in a topic or just due to the time constraints. However, these were included in the survey nonetheless just because of their awesomeness.

NeRF

NeRF-SLAM

NeRF-SLAM [1] proposes a realtime, photometrically accurate scene reconstruction by use of NERFs and Dense Depth inference. The optical flow model by Droid-SLAM, run on a pair of consecutive frames produces, input to dense BA. Then the marginal depth uncertaintes are used in the radiance field training, i.e. the depth loss entries are weighted by the depth measurement information.

FVLoc-NERF

FVLoc-NERF [2] differentiates from competition in that the RGB images are the sole input for localisation, whereas iNeRF, LocNeRF require multimodal sensors to establish the initial pose. NeRF uses a photometric loss iteratively to solve an inverse problem, obtaining the pose from the view. An optimisation here is such that the inverse NeRF iterations are interdispersed with PnP iterations, where instead of the costly neural network inference the relative tranform problem is solved between couple of images.

Differentiable Physics Simulation of Dynamics-Augmented Neural Objects

Authors [3] infer object-level physical properties from NeRF-described objects: mass, center of mass, inertia. These quantities are then used within the DOJO simulator. This isn't trivial, because contact forces arise from interaction between surfaces, but the NERFs don't provide explicit surfaces, but differential probabilities of occupancy instead. The probabilistic model of contact is proposed along the MC procedure for computing the contact forces under this model. Importantly, an Object-Centric Neural Scattering Function is trained to represent the object in a body coordinate frame, contrary to default implementations of NeRF, which are hard to re-position. The idea behind the contact model is the interpenetration between bodies in contact. This overlap is estimated as an intergral of a product of bodies density fields. The repulsive force of bodies is proportional to the interpenetration. Future work will address multi-point contacts, as well as articulated and non-rigid bodies.

Manipulation

A Stable Adaptive Extended Kalman Filter for Estimating Robot Manipulators Link Velocity and Acceleration

How can a serial manipulator benefit from having multiple IMUs installed at its joints? Authors [4] propose an observer (adaptive EKF) that estimates link velocities and accelerations, while being robust to changes in the measurement noise. The filter is adaptive in that the prediction step and the correction step covariance matrices are updated. A sliding window of measurement innovations is maintained, and the avarage of innovations outer products is used for covariance estimation. Link encoders are used only for obtaining the robot trajectory ground truth. This contribution was tested online on the robot using the 1 kHz EtherCAT connection between IMUs and a control PC (real-time Ubuntu).

Workspace Force/Acceleration Disturbance Observer for Precise and Safe Motion Control

An impedance controller is proposed [5] that is augmented by the disturbance observer. A better performance is demonstrated in comparison with conventional impedance control methods. Authors show 1) how vanilla impedance control can result in tracking error due to friction and modeling uncertainty, 2) how introducing the conventional WS-DOB (workspace disturbance observer) can help to eliminate the tracking error, but how it cancels the external force, which may result in the unsafe control, 3) how the proposed WFADOB observer considers external force a control objective, thus maintaining the interaction force to the set desired impedance. The experiments are conducted on the robot with a 1 kHz controller, the F/T sensor mounted at the end-effector is used for evaluation.

Orientation Control with Variable Stiffness Dynamical Systems

This work [6] is concerned with combining motion generation with control, to produce the reactive behaviors, otherwise infeasible with the traditional time-indexed tracking controllers. This is a follow-up publication that extends their initial VSDS framework by supporting the end-effector orientation control. In experiments the robot successfully fullfills a contact-rich task (the cutting of a raw clay brick). The rotational velocity tracking ability is a part of future work.

Null-Space Compliance Variation for Safe Human-Robot Collaboration in Redundant Manipulators Using Safety Control Barrier Functions

According to authors [7], impedance control already effectively reduces the energy transfer during an accidental collision in the colloborative scenario, but the Safety Control Barrier functions can improve the safety further. The null space projection is employed when an obstacle avoidance subtask doesn't involve the cartesian motion of the robot end-effector. The formulation is presented where the task-space control is decoupled from the null-space control. The proof of stability is provided.

I2mpedance - a Passivity Based Integrative Impedance Controller for Precise and Compliant Manipulation and Interaction

Similarly to [5], this work [8] proposes an improvement over the standard impedance control, which aims to decrease the position error while remaining compliant. This is achieved by introducing the integral term. The passivity is ensured by the virtual energy tank. Besides, the force increase due the integration term is limited in the vincinity of the target pose. The robot is controlled at 100 Hz by the ros node running on a PC.

Single-level differentiable contact simulation

The work [9] combines optimization-based dynamics with optimization-based collision detection. These methods were used before for the global planning in contact-rich manipulation. This contribution supports a rich set of shape primitives, while ensuring the contact point uniqueness with the single-level formulation.

Visuo-Tactile Feedback-Based Robot Manipulation for Object Packing.

A manipulation strategy is implementeed [10] by two neural networks: the first is to infer an affordance map, which allows to predict graspability and pushability. Then, the second network, trained via reinforcement (DQN) implements the motion planning.

Step Toward Deploying the Torque-Controlled Robot TALOS on Industrial Operations

Authors [11] employ the whole-body model prediction control for the TALOS robot. The contribution is the differential dynamic programming solver with cost shaping. The proposition is being made that the information necessary for motion planning must be injected into the system, instead of making the use of heuritics. An efficient implementation allows controller to run at 100Hz.

Miscellaneous

UMIRobot: An Open-{Software, Hardware} Low-Cost Robotic Manipulator for Education

Robotics popularization initiative [12], consisting of a plaform, Arduino, PC software, and a set of lectures.

Bi-Manual Robot Shoe Lacing

ABB Yu Mi robot succeeds [13] in lacing a sneaker. Navigation system enables eyelets tracking, different lacing patters are considered. The manipulation process goes through a set of subtasks.

A Unified Trajectory Generation Algorithm for Dynamic Dexterous Manipulation

The single-arm regrasping and the dual-arm handover, both in a very dynamic manner were shown on the embodied robot [14]. The object trajectory generation is posed as an optimal control problem. The robot trajectory is solved with extra contraints in a dedicated procedure. This trajectory is tracked by the admittance controller.

Bistable Tensegrity Robot with Jumping Repeatability Based on Rigid Plate-Shaped Compressors

A very impressive tensegrity robot that jumps [15].

Athletic Mobile Manipulator System for Robotic Wheelchair Tennis

Authors [16] aspire to promote more research in human-scale robot athletics.

EELS: Towards Autonomous Mobility in Extreme Terrain with a Versatile Snake Robot with Resilience to Exteroception Failures

A snake robot [17] with the multi-layer motion control system, comprised of independent exteroceptive and proprioceptive feedback loops.

Conclusion

Many ideas are left unpresented, and many works that caught author's eye were passed over due to the limited time. But, all in all, that is a one way to capture what were happening on IROS this year. Author aspires to make it in person to the future editions of IROS. So, I will see you then.

The figures in this report are courtesy of the corresponding researchers.

References

  1. Rosinol, A., Leonard, J.J. and Carlone, L., 2022. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641.
  2. G. Wenzhi, B. Haiyang, M. Yuanqu, L. Jia and C. Lijun, FVLoc-NeRF : Fast Vision-Only Localization within Neural Radiation Field, IROS
  3. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z. and Schwager, M., 2023. Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters, 8(5), pp.2780-2787.
  4. Birjandi S., Khurana H., Billard A. G., Haddadin S., 2023. A Stable Adaptive Extended Kalman Filter for Estimating Robot Manipulators Link Velocity and Acceleration, IROS
  5. Han W., Yun W., Oh S., 2023. Workspace Force/Acceleration Disturbance Observer for Precise and Safe Motion Control, IROS
  6. Michel, Y., Saveriano, M., Abu-Dakka, F.J. and Lee, D., 2023. Orientation Control with Variable Stiffness Dynamical Systems. arXiv preprint arXiv:2309.15624.
  7. Ducaju, J., Olofsson B., Robertsson A., Johansson R., 2023. Null-Space Compliance Variation for Safe Human-Robot Collaboration in Redundant Manipulators Using Safety Control Barrier Functions, IROS
  8. Voigt F., Naceri A., Haddadin S., 2023. I2mpedance - a Passivity Based Integrative Impedance Controller for Precise and Compliant Manipulation and Interaction, IROS
  9. Le Cleac'h, S., Schwager, M., Manchester, Z., Sindhwani, V., Florence, P. and Singh, S., 2023. Single-level differentiable contact simulation. IEEE Robotics and Automation Letters.
  10. Liang, W., Fang, F., Acar, C., Toh, W.Q., Sun, Y., Xu, Q. and Wu, Y., 2023. Visuo-Tactile Feedback-Based Robot Manipulation for Object Packing. IEEE Robotics and Automation Letters, 8(2), pp.1151-1158.
  11. Perrot, C. and Stasse, O., 2023, October. Step Toward Deploying the Torque-Controlled Robot TALOS on Industrial Operations. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (pp. 10405-10411). IEEE.
  12. Marinho, M.M., Lin, H.C. and Zhao, J., 2023. UMIRobot: An Open-{Software, Hardware} Low-Cost Robotic Manipulator for Education. arXiv preprint arXiv:2301.06668.
  13. Luo, H., Demiris Y., 2023. Bi-Manual Robot Shoe Lacing, IROS
  14. Zhou, C., Gao, W., Lu, W., Long, Y., Yang, S., Zhao, L., Huang, B. and Zheng, Y., 2023. A Unified Trajectory Generation Algorithm for Dynamic Dexterous Manipulation, IROS
  15. Shimura K., Iwamoto N., and Umedachi T., 2023. Bistable Tensegrity Robot with Jumping Repeatability Based on Rigid Plate-Shaped Compressors, IROS
  16. Zaidi, Z., Martin, D., Belles, N., Zakharov, V., Krishna, A., Lee, K.M., Wagstaff, P., Naik, S., Sklar, M., Choi, S. and Kakehi, Y., 2023. Athletic mobile manipulator system for robotic wheelchair tennis. IEEE Robotics and Automation Letters, 8(4), pp.2245-2252.
  17. {Thakker, R., Paton, M., Strub, M. P., Swan, M., Daddi, G. et al., 2023. EELS: Towards Autonomous Mobility in Extreme Terrain with a Versatile Snake Robot with Resilience to Exteroception Failures