TMmichi
commented
4 years ago sim-to-real robot learning from pixels with progressive nets (2016) Rusu et al. https://arxiv.org/pdf/1610.04286.pdf
Contrib: Leveraging on Progressive Nets, trained model from simulation where fast rendering and multi-threaded training environment is available, can successfully provide transfer learning technique without being restricted to the change of model structure (capacity) or modality of input data, and also can reduce the inter-domain variance of simulation and real.
Used input data: 1st column - 64x64x3, 2nd - 64x64x3 + 9 angles + 9 velocities, 3rd - 9 + 9 Used policy appx: progressive nets <Conv + LSTM/FC> Used policy optimization method: A3C (simulation), A2C (real)
Key points:
- Adversarial training of NN can be used to reduce the sensitivity of a trained network.
- Direct low-level control of each joints with policy network without any filtering can also provide stable behaviour. https://youtu.be/qzMTPzbPV0c
Future works: Implementing IRL/IL may results decreasing in training time.
Sim-to-Real via Sim-to-Sim: Data-efficient Robotic Grasping via Randomized-to-Canonical Adaptation Networks (2019) http://openaccess.thecvf.com/content_CVPR_2019/papers/James_Sim-To-Real_via_Sim-To-Sim_Data-Efficient_Robotic_Grasping_via_Randomized-To-Canonical_Adaptation_Networks_CVPR_2019_paper.pdf