Q-adaptive guidance
An adaptive method for inference guidance of VLA models. It formulates the choice of guidance weight as a Markov decision process and trains a small critic network with accurately measured action-value functions to dynamically adjust the strength of inference influence on actions.
Timeline (last 90 days)
The paper proposes the Q-adaptive guidance method, which models the choice of guidance weights as a Markov decision process and fits it with an exactly measurable action value function.
Q-adaptive guidance, by freezing the expert policy and fixing the noise, applies a weight at a certain step and uses the current controller to complete the remaining planning, thereby exactly measuring the true value of that weight.
On the Waymo hidden test set, frozen Alpamayo 1.5 combined with rule-based camera and vehicle transforms plus Q-adaptive guidance achieves a new state of the art among published methods, with a scoring feedback score of 8.135159, compared to the previous best published method's 8.079.
The paper has not yet disclosed the failure modes of weight adjustment in extreme scenarios, nor the generalization performance of the critic in out-of-distribution scenarios.
All Facts (4)
The paper proposes the Q-adaptive guidance method, which models the choice of guidance weights as a Markov decision process and fits it with an exactly measurable action value function.
50%UnverifiedQ-adaptive guidance, by freezing the expert policy and fixing the noise, applies a weight at a certain step and uses the current controller to complete the remaining planning, thereby exactly measuring the true value of that weight.
50%UnverifiedOn the Waymo hidden test set, frozen Alpamayo 1.5 combined with rule-based camera and vehicle transforms plus Q-adaptive guidance achieves a new state of the art among published methods, with a scoring feedback score of 8.135159, compared to the previous best published method's 8.079.
50%UnverifiedThe paper has not yet disclosed the failure modes of weight adjustment in extreme scenarios, nor the generalization performance of the critic in out-of-distribution scenarios.
50%