Abst
- Transfer of procedural knowledge
LIBERO, a novel benchmark
- how to efficiently transfer declarative knowledge
- how to design effective policy architectures
- effective algorithms for LLDM b(lifelong learning in decision making)
- the robustness of a lifelong learner
- the effect of model pre training for LLDM
Intro
Contribution
- Transformer-based architecture is better at abastracting temporal information
- LL algorithms are effective at preventing forgetting
- semantically-rich tasks descriptions no better than those of the task IDs
- Basic supervised pretraining of a large-scale offline dataset → negative impact in LLDM
Background
ロボット学習は有限マルコフ連鎖
LILM
Research topics in LLDM
- Transfer of different types of knowledge
- 他の知識を転用する的な感じかな？
- Neural architecture design
- Lifelong leaning algorithm design
- Robustness of task ordering
- usage of pre-trained model
Procedural Generation of Tasks
- Extract behavioral templates from language annotations
- Specify an initial object distribution
- Specify task goals using a propositional formula
Behavioral templates and instruction generation
Initial state distribution
Task suites
- LIBERO-spatial
- LIBERO-object
- LIBERO-goal
- LIBERO-100
Resnet-RNN, Resnet-T, ViT-T: Vision language policy network
Experiments
- Distribution shifts
Conclusion
- a new benchmark in the robot manipulation domain
- 130 standard tasks
Future direction
- how to design a better neural architecture
- how to design a better algorithm
- how to use pretraining
Top comments (0)