Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/dtch1997/awesome-skill-learning

Collection of resources on skill learning methods
https://github.com/dtch1997/awesome-skill-learning

List: awesome-skill-learning

Last synced: 16 days ago
JSON representation

Collection of resources on skill learning methods

Awesome Lists containing this project

README 

        
# Awesome Skill Learning

  - [What is Skill Learning](#what-is-skill-learning)

  - [Benchmarks](#benchmarks)

  - [Methods](#methods)

    - [Skill Representation](#skill-representation)

    - [Skill Learning](#skill-learning)

    - [Skill Discovery](#skill-discovery)

    - [Transferring Skills to Tasks](#transferring-skills-to-tasks)

    - [Classical Robotics](#classical-robotics)



Collection of resources on skill learning.



## What is Skill Learning



[Blogpost on Skill Learning (WIP)](https://daniel-ch-tan.github.io/blog/2022/skill-learning/)



Tl;dr Learning reusable skills for downstream tasks.  



## Benchmarks



Challenging control environments in which skill learning would be useful



Mujoco:

- Multi-agent soccer environment [code](https://github.com/deepmind/dm_control/tree/main/dm_control/locomotion/soccer)

- CMU motion-capture imitation environment [code](https://github.com/deepmind/dm_control/tree/main/dm_control/locomotion/mocap)



IsaacGym:

- Humanoid adversarial motion prior [code](https://github.com/NVIDIA-Omniverse/IsaacGymEnvs/blob/main/isaacgymenvs/tasks/humanoid_amp.py)

- Humanoid adversarial skill embedding [code](https://github.com/nv-tlabs/ASE)



Embodied AI:

- RoboThor, AI2Thor HabitatAI

- BEHAVIOR-1K: A Benchmark for Embodied AI with 1,000 Everyday Activities and Realistic Simulation [paper](https://openreview.net/pdf?id=_8DoIe8G3t)



Games

- Starcraft Multi-Agent Challenge [paper](https://arxiv.org/abs/1902.04043), [code](https://github.com/oxwhirl/smac), [v2](https://github.com/oxwhirl/smacv2)

- MicroRTS-Py [paper](https://arxiv.org/abs/2105.13807), [code](https://github.com/Farama-Foundation/MicroRTS-Py)



## Datasets



- [CMU Graphics Lab Motion Capture Database](http://mocap.cs.cmu.edu/), diverse motion capture data of humans 

- [AcinoSet](https://github.com/African-Robotics-Unit/AcinoSet), motion capture data of cheetahs 



## Methods



### Skill Representation



- [AAAI 1999] Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning: [paper](https://people.cs.umass.edu/~barto/courses/cs687/Sutton-Precup-Singh-AIJ99.pdf)

- [ICLR 2018] Learning an Embedding Space for Transferable Robot Skills [paper](https://openreview.net/pdf?id=rk07ZXZRb)



### Skill Learning 

Learning to execute known skills. 



Complex motor skills for high-DoF robots through imitation, e.g. motion-capture or video. 

- [Arxiv 2019] Hierarchical Visuomotor Control of Humanoids [paper](https://arxiv.org/pdf/1811.09656.pdf)

- [RSS 2020] Learning Agile Robotic Locomotion Skills by Imitating Animals [paper](https://arxiv.org/abs/2004.00784), [github](https://github.com/erwincoumans/motion_imitation)

- [PMLR 2020] CoMic: Complementary Task Learning & Mimicry for Reusable Skills [paper](https://proceedings.mlr.press/v119/hasenclever20a.html)

- [Arxiv 2022] Learning Agile Skills via Adversarial Imitation of Rough Partial Demonstrations [paper](https://arxiv.org/abs/2206.11693), [video](https://www.youtube.com/playlist?list=PLhqs0Oka9VRFrKb9djmEBU-NyewCKHfGP)

- Temporal Logic Imitation: Learning Plan-Satisficing Motion Policies from Demonstrations [paper](https://openreview.net/pdf?id=ndYsaoyzCWv)

- Watch and Match: Supercharging Imitation with Regularized Optimal Transport [paper](https://openreview.net/pdf?id=ZUtgUA0Fuwd)

- Adversarial Skill Networks: Unsupervised Robot Skill Learning from Video [paper](https://ieeexplore.ieee.org/document/9196582)  

- Learning Semantics-Aware Locomotion Skills from Human Demonstration [paper](https://openreview.net/forum?id=JtK7F6D8t-3)

- RoboTube: Learning Household Manipulation from Human Videos with Simulated Twin Environments [paper](https://openreview.net/forum?id=VD0nXUG5Qk)

- Learning and Retrieval from Prior Data for Skill-based Imitation Learning https://arxiv.org/pdf/2210.11435.pdf



### Skill Discovery

Learning to discover unknown skills. 



General methods based on e.g unsupervised RL objectives

- Quality-Diversity [website](https://quality-diversity.github.io/)

- [Arxiv 2018] Diversity is All You Need: Learning Skills without a Reward Function: [paper](https://arxiv.org/abs/1802.06070), [website](https://sites.google.com/view/diayn/), [code](https://github.com/ben-eysenbach/sac)

- [Arxiv 2020] Dynamics-Aware Unsupervised Discovery of Skills: [paper](https://arxiv.org/abs/1907.01657), [code](https://github.com/google-research/dads)

  - Follow-up work: Emergent Real-World Robotic Skills via Unsupervised Off-Policy Reinforcement Learning [paper](https://arxiv.org/pdf/2004.12974.pdf)

- Learning Temporally Extended Skills in Continuous Domains as Symbolic Actions for Planning [paper](https://openreview.net/pdf?id=t-IO7wCaNgH)

- [ICLR 2023] Efficient Planning in a Compact Latent Action Space: [paper](https://arxiv.org/abs/2208.10291), [code](https://github.com/ZhengyaoJiang/latentplan)



Methods focusing on motor skills for robots

- [CASE 2021] Building Skill Learning Systems for Robotics [paper](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9551562)



Large-scale multi-task training

- Robotics Transformer [blogpost](https://ai.googleblog.com/2022/12/rt-1-robotics-transformer-for-real.html)

- MT-Opt: Continuous Multi-Task Robotic Reinforcement Learning at Scale [paper](https://arxiv.org/abs/2104.08212)



### Transferring Skills to Tasks

Enabling transfer of pre-trained skills to downstream tasks. 



Planning with skills

- World Model as a Graph: Learning Latent Landmarks for Planning [paper](https://arxiv.org/abs/2011.12491)

- Search on the Replay Buffer: Bridging Planning and Reinforcement Learning [paper](https://arxiv.org/abs/1906.05253)

- Skill-based Model-based Reinforcement Learning [paper](https://arxiv.org/abs/2207.07560)



Policy distillation 

- [ICLR 2019] Neural probabilistic motor primitives for humanoid control [paper](https://arxiv.org/abs/1811.11711), [video](https://www.youtube.com/watch?v=CaDEf-QcKwA)



Hierarchical control (e.g. options)

- [Arxiv 2021] From Motor Control to Team Play in Simulated Humanoid Football [paper](https://arxiv.org/abs/2105.12196), [video](https://youtu.be/KHMwq9pv7mg)

- [ACM SIGGRAPH 2022] ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters: [paper](https://arxiv.org/abs/2205.01906), [website](https://xbpeng.github.io/projects/ASE/index.html), [code](https://github.com/nv-tlabs/ASE)

- [Arxiv 2022] Leveraging Approximate Symbolic Models for Reinforcement Learning via Skill Diversity [paper](https://arxiv.org/abs/2202.02886)

- A Deep Hierarchical Approach to Lifelong Learning in Minecraft [paper](https://arxiv.org/abs/1604.07255)

- Residual Skill Policies: Learning an Adaptable Skill-based Action Space for Reinforcement Learning for Robotics [paper](https://openreview.net/pdf?id=0nb97NQypbK)

- PADL: Language-Directed Physics-Based Character Control [paper](https://arxiv.org/pdf/2301.13868.pdf)



Skills as policy priors (similar to student-teacher)

- [RSS 2020] Composable Energy Policies for Reactive Motion Generation and Reinforcement Learning [paper](http://www.roboticsproceedings.org/rss17/p052.pdf)

- [ICRA 2021] Model Predictive Actor-Critic: Accelerating Robot Skill Acquisition with Deep Reinforcement Learning [paper](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9561298), [code](https://github.com/dnandha/mopac)



Program synthesis

- [Arxiv 2021] Learning to Synthesize Programs as Interpretable and Generalizable Policies [paper](https://arxiv.org/abs/2108.13643), [website](https://clvrai.github.io/leaps/)

- [Arxiv 2023] Hierarchical Programmatic Reinforcement Learning via Learning to Compose Programs [paper](https://arxiv.org/abs/2301.12950)



### Classical Robotics

Classical robotics heuristics for implementing high-quality 'skills'.



DMPs.

- [IEEETrans 2013] Dynamical Movement Primitives: Learning Attractor Models for Motor Behaviors [paper](https://ieeexplore.ieee.org/document/6797340)

- [Arxiv 2021] Dynamic Movement Primitives in Robotics: A Tutorial Survey: [paper](https://arxiv.org/pdf/2102.03861.pdf)



Hybrid zero dynamics

- [IEEETrans 2015] Hybrid Zero Dynamics of Planar Bipedal Walking [paper](https://web.eecs.umich.edu/~grizzle/papers/Grizzle_Westervelt_HZD_IsidoriFest.pdf)

- [IJRR 2019] Combining trajectory optimization, supervised machine learning, and model structure for mitigating the curse of dimensionality in the control of bipedal robots: [paper](https://journals.sagepub.com/doi/pdf/10.1177/0278364919859425)

  - This paper introduces generalized HZD which uses ML to interpolate HZD trajectories to obtain a control manifold. 



Other concepts: Locomotion gaits, Raibert heuristic.