GVC: Why Action Modeling Helps Robot Reinforcement Learning.

Generative Value Classifier • Alexander Swerdlow, Perry Dong, Chelsea Finn • Stanford University • arXiv • 2026.

Cold Open.

Imagine training a robot to assemble a stand and hang a wrench. Several movements look reasonable. One preserves the alignment; another creates a problem the robot will discover much later. The action generator needs a repertoire. The critic needs judgment.

Those jobs consume the same experience, yet receive different supervision. A conventional scalar critic learns toward a return target. An action generator learns the structure of movements themselves. Distributional reinforcement learning enriches the critic’s target by predicting possible returns rather than only their mean.

GVC asks a different question: can the actions themselves help train the critic? Its central move is to turn return-conditioned action modeling into a way of scoring actions.

That puts the drama in an unusual place. The new ingredient is not a better way to propose motions. It is a different reason for the critic to understand them.

The Idea.

GVC makes future return the class label. It divides returns into bins and builds two models: an action model conditioned on state and return, and a return prior conditioned on state alone. For a candidate action, combine its score under each bin with that bin’s prior probability. Normalize, then average the bin values to recover Q.

The precedent comes from generative classification. In Your Diffusion Model is Secretly a Zero-Shot Classifier, Alexander Li and colleagues use an image generator for recognition. They compare how well the model explains an image under different labels, using denoising errors to approximate conditional likelihoods. A model trained to generate images can therefore answer a classification question.

Think about what that inversion buys you. Suppose an image contains features compatible with several categories. The useful question is not simply whether the image looks plausible. It is which category best explains those features. Applied to control, the analogous question becomes: which possible outcome best explains this particular movement?.

Here is a hypothetical example. A small wrist rotation might occur in both successful and unsuccessful trajectories. Its overall frequency tells you little. What matters for judgment is whether that rotation is especially characteristic of good outcomes in the current situation. The prior supplies the starting expectations; the action supplies additional evidence.

GVC implements its action scores through flow-matching errors—likelihood proxies—with a shared trunk and learned return-bin embeddings. The candidate action remains central: its noised version enters the network. The change is prediction direction, not an action-blind critic.

Training combines a return-weighted action-generation loss, prediction of the return prior, and squared-error anchoring of Q to a bootstrapped Bellman target. The return labels evolve through temporal-difference learning. This is a hybrid generative-discriminative objective.

That pairing addresses two different engineering goals. You want the representation to preserve detailed structure in motor commands. You also want its preferences to track long-term consequences. Imagine two equally smooth trajectories, one successful and one unsuccessful. Modeling smoothness is useful; deciding between them requires outcome information. The training design gives those jobs a common representation.

A separate behavior-cloned flow actor supplies candidates; GVC selects the highest-valued one. This proposal-and-ranking arrangement has a useful precedent in the best-of-N baseline introduced with Q-chunking: an expressive behavior model supplies options, and value estimates determine the choice.

Decision Transformer offers another contrast. It learns trajectory-conditioned action prediction, with desired return acting as an instruction for generation. That treats return as a request for a particular level of behavior. GVC instead uses return conditioning to judge proposed actions.

This also explains the attraction for robot foundation models. Pi-zero combines a pretrained vision-language model with flow-matching action generation and learns from heterogeneous robot data. Action modeling is already an established interface for large robot policies. The research opportunity is to make that interface useful for judgment as well as generation.

The Result.

The clearest result is simulated Tool Hang: GVC reaches roughly 85% success within half a million environment steps; Q-chunking reaches roughly 50% after twice that interaction budget.

Online training mixes equal parts human demonstrations and replay, with the critic evaluating short action chunks. The online comparison uses three seeds; the OGBench experiments and main ablations use eight.

The demonstration source matters. Robomimic was developed around human manipulation data, including high-precision insertion and assembly. Its datasets distinguish proficient operators from mixed-proficiency demonstrators. This gives the learning problem behavioral structure to exploit, rather than reducing it to exploration from arbitrary motor noise.

Q-chunking is a meaningful comparator because it already addresses long-horizon learning. It operates over action sequences, encourages temporally coherent exploration, and enables multi-step value backups. GVC adopts that training recipe and changes the critic. The comparison therefore starts with a capable mechanism for producing sustained behavior.

The offline evaluation covers 35 OGBench tasks. A separate six-task offline-to-online evaluation gives GVC the highest final average success. BFN supplies the close control: the same flow actor and best-of-N selection, with a regression critic.

OGBench makes this an interesting test of decision quality. Its manipulation datasets include scripted play: picking and placing blocks, interacting with drawers and buttons, and pressing puzzle buttons in arbitrary sequences. Evaluation asks the agent to combine those behaviors into desired configurations. The puzzle domain adds combinatorial structure: pressing a button changes neighboring buttons, so locally plausible behavior must support a coherent sequence.

That is why I care about the benchmark beyond its task count. It asks whether familiar movements can be assembled into useful behavior. For a critic-learning paper, that is a more revealing question than whether the policy can reproduce an isolated demonstration.

The strongest distributional comparison is Value Flows, which models the return distribution with a flow model and uses return uncertainty to weight learning. It provides a distinct alternative: improve the model of outcomes themselves.

The GVC ablations support the action-supervision explanation. Reducing the action-modeling loss hurts Tool Hang performance. Training that loss uniformly across return bins recovers part of the benefit; return-aware weighting contributes further improvement.

That intervention is especially informative. Keep the architecture, change what it must learn, and behavior changes. The result directs attention toward the training signal—not simply the presence of a generative network.

The Biggest Weakness.

The transition from benchmark critics to pretrained robot policies is untested. The reported models are state-input MLPs evaluated in simulation, not adaptations of a pretrained vision-language-action policy.

That matters because action supervision can serve different purposes at different stages of learning.

Consider a network starting from scratch. Learning to predict actions might teach it basic regularities about motion, grasping, and the relationships between states and commands. Now consider a pretrained policy such as pi-zero, which has already learned action generation across diverse robotic experience. Much of that behavioral structure is already part of the starting point.

Two hypotheses follow.

One is that action prediction primarily helps construct a useful representation when the critic has little prior knowledge. Another is that organizing an existing action representation around possible returns provides additional value even after extensive pretraining. Those hypotheses can both explain the benchmark motivation, yet imply different benefits for foundation-model fine-tuning.

This is an experimental distinction, not a semantic argument about what counts as generative.

The decisive follow-up would use the same pretrained backbone, action candidates, downstream data, and interaction budget for a conventional critic and a GVC-style critic. A corresponding from-scratch comparison would reveal how much each approach gains specifically from initialization.

The outcome to examine is the additional improvement attributable to pretraining. Does the generative formulation make existing motor knowledge more useful for estimating value? Does it need fewer downstream interactions to turn that knowledge into better choices?.

I would also want the evaluation to preserve the intended observation interface. Testing a pretrained visuomotor system should involve judgments from its visual inputs, rather than substituting privileged simulator state.

That experiment would connect the proposed architectural compatibility to the practical claim: learning better value estimates from a robot policy that already knows how to move.

Verdict.

GVC deserves attention from researchers building sparse-reward robot RL, especially those already working with expressive action generators and demonstration data.

The most useful contribution is a change in how we think about critic training. Treat the critic as a representation learner with several things worth understanding, rather than as the final numerical stage of an otherwise sophisticated policy system.

The surrounding literature makes the distinction clearer. Q-chunking changes the temporal unit of decision-making. Value Flows enriches the model of future returns. These are separate ways to improve value-based control: change the decision horizon, or change the description of outcomes. GVC adds another axis—what information in the action data should shape the critic’s representation.

For practitioners, I would frame the opportunity around the separation between proposing and choosing. When an action generator already offers plausible alternatives, improving the judge becomes a focused research problem. You can ask whether the system prefers the right option without simultaneously redesigning the entire mechanism that produces options.

That separation also makes experimental results easier to interpret. A stronger proposal distribution changes what the system can consider. A stronger ranking mechanism changes what it chooses from those possibilities. Both matter; keeping the distinction explicit makes a critic result more informative.

The broader lesson is to examine each logged transition as a collection of potential learning targets. The reward describes the task’s consequences. The action contains a detailed example of behavior. A good learning objective can make those signals cooperate.

That is the intellectual appeal of GVC: generation and evaluation become two uses of related behavioral structure. The question shifts from “How expressive should the value head be?” toward “What should the value learner understand?”.

The takeaway: give the critic action structure to learn, as well as values to fit.

Outro.

The next development to watch is whether action-supervised critics become a reusable ingredient across different reinforcement-learning systems.

Return to the robot holding the wrench. It needs plausible movements, and it needs a reason to prefer one movement over another. Those are distinct demands on the learner.

GVC makes their relationship worth investigating more carefully. When an experiment stalls, the interesting question may be what the critic is being taught to understand—not only how many layers it has or how many actions the policy can sample.

That is a useful question to carry into the next robot-learning experiment.