RESUMO
Contact-rich robotic manipulation tasks such as assembly are widely studied due to their close relevance with social and manufacturing industries. Although the task is highly related to vision and force, current methods lack a unified mechanism to effectively fuse the two sensors. We consider coordinating multimodality from perception to control and propose a vision-force curriculum policy learning scheme to effectively fuse the features and generate policy. Experiments in simulations indicate the priorities of our method, which could insert pegs with 0.1 mm clearance. Furthermore, the system is generalizable to various initial configurations and unseen shapes, and it can be robustly transferred from simulation to reality without fine-tuning, showing the effectiveness and generalization of our proposed method. The experiment videos and code will be available at https://sites.google.com/view/vf-assembly.
RESUMO
PURPOSE: A high-precision rotated elliptical beam profiling method based on pixel ion chamber is proposed in this paper. This method aims to improve the accuracy by modeling the transverse profile of rotated beam as an ellipse with additional correlation coefficient and eliminating the fitting error due to the volume averaging effect of pixel ion chamber. METHODS: In pencil beam scanning (PBS) proton therapy systems, the transverse beam profile model is generally represented as a standard Gaussian distribution. Considering the elliptical spots, two-dimensional (2D) joint Gaussian distribution characterized with the correlation coefficient ρ is adopted in this study. Gaussian-type particle distribution with white noise was generated and processed in MATLAB to simulate the secondary particle collection in the pixel ion chamber. The simulated pixel ion chamber is a commercially available ion chamber which consists of 12 × 12 small square pixels (3.75 × 3.75 mm2 ) with a 0.05 mm interval. The simulated signals were preprocessed by filtering with the noise threshold and extracting the maximum simply connected domain (MSCD) of the signal. Then, five geometric parameters that identify the transverse beam profiles were fitted under different signal-to-noise ratio (SNR) conditions: the center of the beam (x0 , y0 ), the spot size (σmajor , σminor ), and the rotation angle θ formed between the major axes of elliptical spot and the x axes of the ion chamber. First, the simulated signals were preprocessed by filtering with the noise threshold and extracting the MSCD of the signal. Second, a rectification curve of systematic error in fitted spot size versus the prescribed spot size was used to predict the systematic error due to the volume averaging effect. Finally, the effects of fitting errors on therapeutic dose were evaluated in terms of gamma index and relative dose difference. RESULTS: When the SNR is not less than 20 dB, the relative fitting error of spot size and the absolute fitting error of angle θ are less than 1% and 6.1°, respectively. The fitting error of beam center increases with spot size and will not exceed 0.22 mm when spot size reaches up to 12 mm. At a SNR equal to 20 dB, neither cold nor hot spots were presented in dose distribution calculated with the fitted spot parameters. CONCLUSION: The improved Gaussian fitting algorithm performs well when SNR is not less than 20 dB. This method can effectively distinguish the nominal beam and rotated elliptical beam. An ideal systematic error curve can be predicted and used to correct the fitted spot size, thus eliminating the systematic error due to the volume averaging effect of the pixel ion chamber. The fitting error of spot size cannot be fully corrected, but it is negligible and shows little effect on the overall therapeutic dose.