Ahmet Ercan Tekden

[Ahmet Ercan Tekden @ErcanTekden]

Dynamical system methods are effective for reactive motion generation, but tasks that require different motions from similar states at different stages of execution remain challenging.

We address this with Phase-varying Neural Potential Functions (PNPF).

1/8

[Ahmet Ercan Tekden @ErcanTekden]

Joint work with @dkanou, Aude Billard, and Yasemin Bekiroglu.

Paper: https://t.co/lZUhDrnLJB
Project Page: https://t.co/vxHBNzevkN
Colab: https://t.co/lOveOeqSUQ
8/8

[Ahmet Ercan Tekden @ErcanTekden]

Dynamical system methods are effective for reactive motion generation, but tasks that require different motions from similar states at different stages of execution remain challenging.

We address this with Phase-varying Neural Potential Functions (PNPF).

1/8

[Ahmet Ercan Tekden @ErcanTekden]

PNPF enables reactive motion generation in 6D environments, including periodic motions.

The framework remains robust to perturbations while preserving the demonstrated task behavior.

7/8

[Ahmet Ercan Tekden @ErcanTekden]

Please see our presentation on the Grasping 1 session on Tuesday (Room NT-G2, 10:30-12:00), and join our poster session at 13.30 for more details.

Project Page: https://t.co/UE7DtetbNC

#ICRA2024

[Ahmet Ercan Tekden @ErcanTekden]

#ICRA2024 Objects we interact with and manipulate often share similar parts, such as handles, where similar grasps can be applied.

[Ahmet Ercan Tekden @ErcanTekden]

In our work, we model the local surfaces around the grasp locations given on a category of objects, such as mugs, to transfer grasps to novel objects that share a geometrically similar part, such as bags and bowls.

[Ahmet Ercan Tekden @ErcanTekden]

@ftm_guney Are these problems solved? I feel like even with the best available implementations, outputs between different frames for the same scene are inconsistent. Also, people often have to use heuristics just so that their method doesn't fail (Eg. Figure 3 https://t.co/kvLj2hTSdy)

[Ahmet Ercan Tekden @ErcanTekden]

Our method can utilise 3D shapes via signed distance functions for generating motion trajectories (9 Demonstrations). It generalises to rotation-based scene changes and estimates accurate trajectories with 2 cm precision for novel scenes. (6/6)

[Ahmet Ercan Tekden @ErcanTekden]

Happy to share that our paper “Neural Field Movement Primitives for Joint Modelling of Scenes and Motions” has been accepted to #IROS2023. Joint work with Yasemin Bekiroglu and Marc P. Deisenroth.

Project Page: https://t.co/V2IaIBx0xT
Arxiv: https://t.co/hIQN11OpWU

(1/6)

[Ahmet Ercan Tekden @ErcanTekden]

Our method can model multivalued trajectories by modelling motion as an implicit function of end effector positions and time values (18 Demonstrations). Since test-time optimization is used, distractor objects introduced at inference time do not deteriorate the performance. (5/6)

[Ahmet Ercan Tekden @ErcanTekden]

@ftm_guney I have been using A3000 for a while and I did not have any issues, but I didn't use very deep architectures. However it has only 6 Gb VRAM (I guess this issue always exist with laptop GPU's), which prevented me from using some of the recent Photo-realistic simulators.