What’s Changed
LeRobot v0.6.0 provides a technological innovation to close the robot learning loop. The world model policy (VLA-JEPA, FastWAM, LingBot-VA) has the ability to imagine the future, and six new VLAs (GR00T N1.7, MolmoAct2, EO-1, EVO1, Multitask DiT) have been added. Furthermore, new reward model APIs, Robometer and TOPReward, have been introduced. Additionally, deep sensing, VLM-driven dataset annotation, custom video encoding, cloud training on AWS Jobs, and a more lightweight installation have been implemented. (Source: Hugging Face Blog)
Mechanism Details
The world model policy of LeRobot v0.6.0 uses VLA-JEPA and FastWAM to predict the future. VLA-JEPA is based on a visual language architecture, and LingBot-VA achieves integration of language and action. The reward model Robometer provides a mechanism to evaluate the success of robots, and TOPReward measures the degree of task achievement. These technologies accelerate data loading by up to 2 times and support complex simulation benchmarks. (Source: Hugging Face Blog)
Migration Procedure
To run the MiniMax model on AWS Bedrock, specify MiniMax M2.5 and configure it to meet AWS security and compliance requirements. To build multi-turn RL on SageMaker HyperPod, use AWS CDK to provision infrastructure and upload data to Amazon S3. The training job will be automatically started and triggered by Amazon EventBridge. (Source: AWS Blog, AWS Blog)
Performance Characteristics
In LeRobot v0.6.0, data loading is accelerated by up to 2 times, and simulation benchmarks are executed efficiently. MiniMax M2.5 adopts a hybrid expert (MoE) architecture for agent-native execution, reducing inference costs. SageMaker HyperPod provides serverless training for multi-turn RL, preventing waste of GPU resources. (Source: Hugging Face Blog)
Summary
LeRobot v0.6.0 offers significant improvements in robot learning, including enhanced world model policies, new VLAs, and improved reward models. The migration procedure for running MiniMax models on AWS Bedrock and building multi-turn RL on SageMaker HyperPod is also outlined. The performance characteristics of LeRobot v0.6.0 demonstrate accelerated data loading, efficient simulation benchmarks, and reduced inference costs. (Source: Hugging Face Blog)