Chronos Policy
A robot policy model based on the Diffusion Transformer architecture that explicitly incorporates physical timestamps as inputs, enabling the policy to perceive real time intervals. It supports asynchronous multi-frequency sensor data and improves temporal robustness in robot control tasks.
Timeline (last 90 days)
The core claim of Chronos Policy is to make physical time an explicit input to the policy, letting the model know the acquisition timestamp of each observation and the expected execution timestamp of each action.
Chronos Policy is built on a diffusion transformer architecture.
The researchers explored two methods for introducing physical time into the model: timestamp embeddings and time-aware rotary position encoding (time-aware RoPE).
Experimental results indicate that time-aware RoPE is more robust than timestamp embeddings.
During training, Chronos Policy applies temporal randomization to two things: randomly sampling which observations are used, and randomly deciding when the action chunk begins execution.
Temporal randomization during training allows the policy to learn from asynchronous, multi-frequency data without needing to align all measurements to a fixed time grid.
On dynamic tasks, Chronos Policy delivers clear performance improvements.
On multi-sensor tasks, Chronos Policy performs better.
On standard quasi-static tasks, Chronos Policy maintains nominal performance comparable to the baseline while improving robustness to temporal variations in perception and inference.
When a policy must face real-world temporal uncertainty, it is better to provide time as an explicit input to the model than to let the model implicitly guess the time.
All Facts (10)
The core claim of Chronos Policy is to make physical time an explicit input to the policy, letting the model know the acquisition timestamp of each observation and the expected execution timestamp of each action.
50%UnverifiedChronos Policy is built on a diffusion transformer architecture.
50%UnverifiedThe researchers explored two methods for introducing physical time into the model: timestamp embeddings and time-aware rotary position encoding (time-aware RoPE).
50%UnverifiedExperimental results indicate that time-aware RoPE is more robust than timestamp embeddings.
50%UnverifiedDuring training, Chronos Policy applies temporal randomization to two things: randomly sampling which observations are used, and randomly deciding when the action chunk begins execution.
50%UnverifiedTemporal randomization during training allows the policy to learn from asynchronous, multi-frequency data without needing to align all measurements to a fixed time grid.
50%UnverifiedOn dynamic tasks, Chronos Policy delivers clear performance improvements.
50%UnverifiedOn multi-sensor tasks, Chronos Policy performs better.
50%UnverifiedOn standard quasi-static tasks, Chronos Policy maintains nominal performance comparable to the baseline while improving robustness to temporal variations in perception and inference.
50%UnverifiedWhen a policy must face real-world temporal uncertainty, it is better to provide time as an explicit input to the model than to let the model implicitly guess the time.
50%