π0.5
A vision-language-action (VLA) model that end-to-end integrates large pretrained language and vision models with robot action prediction, designed for the evaluation and generalization study of robot manipulation policies.
Timeline (last 90 days)
Backtrack was evaluated on the LIBERO and MetaWorld benchmarks, covering three VLA models: SmolVLA, π0.5, and π0-FAST.
Flow-matching action experts (such as π0.5) model action generation as a continuous transformation process from noise to actions, producing smooth action sequences through iterative denoising and suiting high-dimensional continuous control; autoregressive methods (such as π0-FAST) discretize actions into tokens and predict them token by token.
REKI achieves 90.4% on the LIBERO-Plus benchmark, surpassing Qwen-RobotManip (89.0%) and π0.5 (85.7%).
On a single H100 with torch.compile, BF16 precision and two 256x256 RGB inputs, REKI produces a complete action chunk in only 21.8 milliseconds, 35.5% faster than π0.5 (33.8 ms).
Actor Labs' excavator experiment deployed Physical Intelligence's π0.5 onto an excavator, combined with large-scale real data from industrial operators to train the policy, and completed a task with only 200 trajectories, exposing the shortcomings of the purely data-driven route: extremely low data efficiency and limited multi-task capability.
The research team conducted tests on two vision-language-action (VLA) models, OpenVLA-OFT and π0.5, covering two benchmark environments, LIBERO-Spatial and LIBERO-Plus.
All Facts (6)
Flow-matching action experts (such as π0.5) model action generation as a continuous transformation process from noise to actions, producing smooth action sequences through iterative denoising and suiting high-dimensional continuous control; autoregressive methods (such as π0-FAST) discretize actions into tokens and predict them token by token.
50%UnverifiedREKI achieves 90.4% on the LIBERO-Plus benchmark, surpassing Qwen-RobotManip (89.0%) and π0.5 (85.7%).
50%UnverifiedOn a single H100 with torch.compile, BF16 precision and two 256x256 RGB inputs, REKI produces a complete action chunk in only 21.8 milliseconds, 35.5% faster than π0.5 (33.8 ms).
50%UnverifiedActor Labs' excavator experiment deployed Physical Intelligence's π0.5 onto an excavator, combined with large-scale real data from industrial operators to train the policy, and completed a task with only 200 trajectories, exposing the shortcomings of the purely data-driven route: extremely low data efficiency and limited multi-task capability.
50%UnverifiedThe research team conducted tests on two vision-language-action (VLA) models, OpenVLA-OFT and π0.5, covering two benchmark environments, LIBERO-Spatial and LIBERO-Plus.
50%UnverifiedBacktrack was evaluated on the LIBERO and MetaWorld benchmarks, covering three VLA models: SmolVLA, π0.5, and π0-FAST.
50%