1. Motivation of The Work

• providing a robust safety guarantee for robot tasks with constant force and contact (typical industrial automation tasks)

• minimize the sim2real gap of contact tactile skill

2. Our Work

• Estimating contact-rich task energy distribution with data-driven method

• Tactile-Morph Skills Framewor = task energy distribution modeling + unified force-impedance control 

  • providing a formal passivity guarantee for contact-rich tasks

3. Baseline Comparison

1.Estimation Method Comparison

We compare our energy distribution backbone with the existing time-series estimation pipeline with State-of-the-art performance, namely MLP  and transformed-based framework TST on our self-collected dataset. The results can be seen below.

Figure 2: The Estimated Energy distribution of MLP, TST, TCN and the ground truth

2. Controller Baseline Comparison

Unified force-impedance control (UFIC) can integrate the advantages of impedance control and force control, ensure the compliance of robot systems, exact force regulation, and passivity of the systems, which is crucial in modern robotic applications.

Table 2: Controller Comparison: Impedance Control, Force Control and Unified Force-Impedance Control

Figure 3. Various Automation Scenarios that require force application and contact at the same time 

4. More test Scenarios

1. Different Geometry

As described in Section 4.2, table 2 in paper, we tested our estimation pipeline’s generalizability in a zero-shot setup on the planar and inclined surfaces, where a significant increase of all errors can be observed in the planar and inclined surfaces, this could lead to either insufficient energy, causing performance decoding, or overestimating energy, resulting in a longer ”reflection time” in handling unexpected contact loss.

So, we extended real robot experiment on surface with different geometry in a few-shot setup. First, we gathered 5% more data on the testing surface and trained two extra epochs. The results can be checked in the table below.

With accurate energy distribution, our pipeline can guarantee safety in contact loss cases under different geometries. We tested the tuned pipeline on the planar surface, where the task energy was correctly estimated and quickly drained in contact loss cases. This ensures a proper handling of environmental uncertainties. The falling distance is even smaller than the 3D curved surface since the trajectory is more straightforward and the end-effector has no velocity heading downwards.

Figure 4: Real Robot Experiment on Planar Surface

Planar Surface Experiment Video

2. Different Friction

we tested our framework on surface friction (we removed the whiteboard folie as in figure below); the proposed pipeline failed to capture the correct energy distribution with MAPE over 20%, for it tends to underestimate the needed energy for the task. 

Such behavior makes sense as the input of the pipeline doesn’t include any surface friction information; however, including visual friction detection will be interesting.

Figure 5: Surface with different friction

5. Data Collection

Some more details about automatic real robot data collection

a) Planar Surface                                                                                                                                  b) Inclined Surface