Ablation Study Clause Samples

Ablation Study. The ablation study results are depicted in the Figure 25, showing reference trajectories (dashed-black lines), closed-loop (solid-red lines), and open-loop (solid-blue lines) actual trajectories. This visual representation illustrates each model's capability in tracking trajectories under different conditions. Table 20 summarizes the trajectory tracking results in both open-loop and closed-loop configurations for seen and unseen scenarios. It provides the root mean squared error (RMSE) and standard deviation (σ) for each model, highlighting the ANODE model's superior performance. Table 20: Ablation and comparison results of trajectory tracking scenarios; -O/-C tokens refer to open-loop and closed-loop results, respectively. Scenario Model RMSE x(mm) RMSE y(mm) RMSE z(mm) σ x(mm) σ y(mm) σ z(mm) Seen-Closed (C) MLP 0.531 0.548 0.570 0.531 0.548 0.869 Seen-Open (O) MLP 0.572 0.563 0.702 0.557 0.543 0.494 Seen-C RNN 0.542 0.537 5.100 0.541 0.537 0.498 Seen-O RNN 0.576 0.546 5.000 0.544 0.540 0.419 Seen-C ANODE 0.105 0.127 0.116 0.105 0.127 0.116 Seen-O ANODE 0.198 0.177 0.122 0.198 0.177 0.121 Unseen-C/O MLP/RNN - - - - - - Unseen-C ANODE 0.256 0.164 0.157 0.256 0.164 0.157 Unseen-O ANODE 5.600 3.100 4.900 4.700 1.800 2.900 This section introduced a new method for modelling the continuous forward kinematic models of soft continuum robots using ANODE. The proposed method only required 25 scattered data points. Additionally, we developed a parallel MPPI-based controller running on a GPU, which effectively handles a non-convex objective function. This design enhances the adaptability and robustness of the learned model, enabling accurate prediction and control of soft continuum robot motion in various new scenarios. Through extensive experimentation, ablation, and comparison studies, our proposed framework (ANODE+MPPI) exhibits superior performance over learning-based approaches like FNN and RNN in unseen-before scenarios. It also slightly outperforms them in seen-before settings. Online Estimation of Articulated Objects with Factor Graphs using Vision and Proprioception

Ablation Study. In this ablation experiment, we aim to verify the importance of Expert Proxy, Expert-specific loss, Expert Assignment mechanism, and Expert Con- fidence. To verify the importance of these com- ponents, we remove them one at a time to eval- uate their impacts in terms of w-Acc. and w-F1. on the IEMOCAP dataset. The ablation experi- ment results are shown in Table 3. We can see that when each of the above components is removed, the model’s scores on the w-F1. and w-Acc. met- rics are reduced to varying degrees. In particular, the effects on the model’s performance from large to small are Expert Proxy, Expert-specific Loss, Expert Assignment, and Expert Confidence. When the Expert Proxy is removed, EAAN-ERC can- not assign the proxy model to represent the unseen evaluator, which biases the final evaluation results. When Expert-specific Loss is removed, although the models in the expert pool can represent the eval- uators, they cannot learn the evaluation experience of the corresponding evaluators, thus the perfor- ▇▇▇▇▇ of the EAAN-ERC decreases when perform- ing evaluator-specific emotional evaluation. When Expert Assignment is removed, all expert models in the expert pool are assigned to participate in emo- tion evaluation, which also brings a certain bias to the final evaluation results. We observe that the Expert Assignment has a relatively small impact on model performance. This may be due to the more comprehensive emotional representation extracted by more expert models, despite being perturbed by irrelevant evaluators, which causes a relatively small reduction in model performance. Finally, when Expert Confidence is removed, the perfor- ▇▇▇▇▇ of the model decreases minimally, which means that calculating confidence will strengthen to a certain extent the overall assessment of agree- upon emotions. In summary, through this ablation experiment, we verified how important these com- ponents are to the model performance.

Ablation Study. The conducted ablation study on the CATER-v1 dataset underlines the critical role of the Neural ODE module in our TiV-ODE framework and its resilience against irregular video inputs. By evaluating scenarios with videos of irregular timesteps and replacing the Neural ODE module with a step-wise transition model, we affirm the necessity of our methodological choices. Quantitative results of the ablation study are presented in Table 13. The quantitative outcomes from this study not only reinforce our method's robustness against irregularities in video data but also highlight the Neural ODE module's effectiveness in enhancing our model's video generation capabilities, distinguishing our approach from prior works. Table 13: TiV-ODE Ablation Study on CATER-v1 Dataset TiV-ODE 0.96 11.98 0.12 TiV-ODE - with irregular timesteps 0.96 13.71 0.13 TiV-ODE - without ODE solver 0.90 31.56 0.28

Ablation Study. Additionally, we conducted an ablation study to identify the best way of integrating the textual hints into our MCL framework. In addition to our MCL+Text method, we also explored the following strategies for injecting particles:

Ablation Study. We carried out an ablation study on two high-precision tasks: peg insertion and kits assembling, to evaluate the contribution of the Dynamic-NODE to NODE-IL's effectiveness and to assess NODE-IL's robustness. The study involved two experimental conditions: (i) removing the Dynamic-NODE, allowing the Control-NODE to interact directly with the simulation environment, and (ii) introducing random Gaussian noise X∼N(0,σ^2) to the state observations from the simulation during the testing phase, with experiments conducted across noise levels σ=[0.1,0.2,0.3,0.4,0.5]. The results detailed in Table 17 and Figure 20 reveal that omitting the Dynamic-NODE and directly engaging the Control-NODE with the simulation decreases performance by 5.5%, highlighting the Dynamic-NODE's significant role in NODE-IL's overall efficacy. Furthermore, the NODE-IL framework exhibited resilience to random noise, maintaining robust control policies at noise levels up to σ=0.

Ablation Study. In order to analyze the effectiveness of the proposed Domain-Invariant Embedding Network, we compare the baseline model with ten different configurations. For our model we use DIE features from different the Ts after L2-normalization as retrieved vectors in testing. The result of each experiment is reported in the each row of Table 5.1.

Ablation Study. An ablation study was conducted to validate the effectiveness of the proposed network design. We tested variations of our model on the OCID-Grasp dataset: one without predicting the object class label and another without generating the instance mask.

Ablation Study. To justify our use of both low-level and hierarchical ▇▇▇▇▇- tic information, we conducted an ablation study. We analyzed three strategies for integrating semantic knowledge into an MCL framework. SMCL uses only semantic cues through the semantic visibility model. HMCL uses semantic hierar- chy, described in Sec. III-G, to initialize the particles only in the rooms corresponding to the observed room category, and then relies solely on the LiDAR information. HSMCL combines both strategies. The ATE was computed only on stable sequences with 100% success rate. As can be seen in Tab. VI, utilizing the two levels of semantic information benefits localization. HSMCL was able to localize stably even on the challenging sequences, where other methods failed. The ATE for HSMCL is on par with the other methods, and Method Hierarchy Semantics Success ATE ATE (# of stable sequences) (# of successful runs) MCL 27% - (0) 0.046/0.20 (15) HMCL C 61% 0.046/0.21 (3) 0.044/0.19 (34) SMCL C 81% 0.055/0.23 (7) 0.066/0.24 (45) HSMCL C C 100% 0.079/0.23 (11) 0.079/0.23 (55) TABLE VII: Runtime for HSMCL, with 10,000 particles. The Yolov5s results are for inference on a single camera. Platform Sem. Visibility Beam-End Yolov5s (640x480) Yolov5s (320x240) NUC10i7FNK 55 ms 24 ms 223 ms 57 ms Dell Precision-3640-Tower 19 ms 14 ms 10 ms 6.8 ms the slightly larger error can be attributed to including more challenging sequences and runs in the computation of the ATE for HSMCL, sequences and runs where other methods failed to localize entirely.