VLA-Precision: Asymmetric Co-Bootstrapping for Efficient Real-World Online RL of Vision-Language-Action Models

Abstract: Pretrained vision-language-action (VLA) models enable broad manipulation but remain unreliable in tasks demanding precision and repeatability. Applying real-world online reinforcement learning (RL) to VLA post-training enables autonomous trial-and-error improvement beyond demonstrations alone, but exposes two bottlenecks: 1) unreliable value signals can induce policy drift; 2) large-VLA overhead constrains throughput and sample efficiency. To address these challenges, we present VLA-Precision, an efficient real-world online RL framework featuring the Asymmetric Co-Bootstrapping (ACoB) algorithm and the ACoB-Stream architecture. Specifically, ACoB establishes asymmetric co-bootstrapping across timescales: early intervention-guided behavioral learning rapidly improves policy performance while enhancing online experience quality. As autonomous experience accumulates, global return propagation and local preference ranking progressively calibrate value estimates, yielding relative action advantages for reference-regularized policy improvement while suppressing drift. To enable ACoB on large VLAs, we develop ACoB-Stream, a closed-loop experience--policy architecture that establishes invariant-state decoupling and on-demand streaming as design principles, delivering up to 10.9$\times$ improvements in throughput and computational efficiency. Extensive evaluations on nine high-precision chemistry tasks across four categories and four robot embodiments show that VLA-Precision achieves 98.3\% mean success rate in 45.8 min/task, with 27.6 s episodes running at 1.2$\times$ and 1.8$\times$ the speeds of VLA and RL baselines. Resources are available at https://vla-precision.github.io.

• The paper introduces VLA-Precision, a framework that combines Asymmetric Co-Bootstrapping (ACoB) and ACoB-Stream for efficient real-world online RL of Vision-Language-Action (VLA) models, achieving a 98.3% success rate across nine tasks and completing them 27.6 seconds on average post-training.
• The key methodologies include Asymmetric Co-Bootstrapping (ACoB) optimization algorithm and ACoB-Stream systems architecture, optimizing for value calibration, and stable policy updates across different training timescales and using task-specific critics for improved learning.
• ACoB-Stream enhances computational efficiency by reusing invariant multimodal prefixes, deduplicating context storage, selective context retrieval, and targeted synchronization of action-expert parameters within actor and learner processes.

Research problem and central contribution

“VLA-Precision: Asymmetric Co-Bootstrapping for Efficient Real-World Online RL of Vision-Language-Action Models” (2609.04355) addresses the difficulty of applying online RL to large VLAs in precision-sensitive physical manipulation. The paper identifies two coupled failure modes. First, value estimates derived from sparse, intervention-contaminated real-world trajectories can be unreliable, and direct $Q$-maximization may induce policy drift away from a competent demonstration-initialized policy. Second, the computational cost of large multimodal models—particularly repeated frozen-prefix inference, context storage, random access, and full-model synchronization—reduces the throughput of the actor–learner loop.

The proposed VLA-Precision framework combines two components. Asymmetric Co-Bootstrapping (ACoB) is the optimization algorithm. It combines rapid behavioral learning from successful executions and human corrections with slower value calibration from global returns and local action preferences. ACoB-Stream is the associated systems architecture. It reuses invariant multimodal prefix states, stores contexts through a deduplicated disk-backed mechanism, retrieves only objective-relevant contexts, and synchronizes only the trainable action-expert subspace.

The framework uses a two-stage post-training protocol. Stage I fully fine-tunes a pretrained $\pi_{0.5}$ VLA using task demonstrations. Stage II freezes the multimodal prefix and the base action expert, introduces LoRA parameters in the action expert, and performs asynchronous real-world online RL. This parameterization constrains online adaptation to a relatively small trainable subspace while retaining a frozen Stage-I reference policy for regularization.

ACoB: asymmetric learning across behavioral and value timescales

The conceptual premise of ACoB is that behavioral and value learning should not be treated symmetrically at the beginning of real-world training. Human interventions provide high-quality corrective actions immediately, whereas reliable long-horizon value estimates require accumulated autonomous experience. ACoB therefore gives behavioral cloning a rapid role in improving the policy and the data distribution, while assigning the critic a progressive calibration role.

The learner maintains an ensemble of task-specific critics using a value–advantage decomposition, $Q(\omega,a)=V(\omega)+A(\omega,a)$. The critic receives two complementary signals. Global return propagation uses TD targets based on the executed action, including actions produced autonomously and actions supplied through intervention. This propagates long-horizon task returns through the trajectory. Local preference ranking addresses a problem that TD learning alone cannot resolve: when a human overwrites a poor policy proposal, the executed correction has a successor state and can be trained with TD, but the overwritten proposal does not. Without an explicit comparison, the failed proposal may remain overvalued.

For intervention-marked transitions, ACoB imposes a margin requiring the corrected action to have higher estimated advantage than the original proposal. The ranking is applied to the action-dependent component rather than the full $Q$ value, which removes the state-only value component from the comparison. This is a technically important design choice: the ranking objective is intended to identify local action improvement at a fixed state, rather than to fit an absolute return scale that may vary across states and critics.

Policy improvement is based on a pessimistic relative advantage rather than absolute value maximization. The current action expert and the frozen reference action expert decode using the same noise realization, reducing stochastic variation in their comparison. The current action is compared against a baseline formed from the reference action and, when applicable, the original intervention proposal. The minimum relative advantage across critics is then used as the improvement signal. Consequently, an update is favored only when all critics support the current action relative to the baseline.

This mechanism is combined with a soft-margin loss, intervention-guided flow-matching behavior cloning, and reference regularization. The behavioral-cloning mask includes all chunks from successful trajectories and only effective human corrections from failed trajectories. Reference regularization acts on body-action coordinates and constrains the online policy toward the Stage-I policy. The complete action-expert loss uses weights $(0.25,0.50,0.25)$ for behavior cloning, relative-advantage improvement, and reference regularization, respectively.

The resulting division of labor is explicit. Behavioral cloning rapidly incorporates reliable corrections and successful behavior; improved behavior produces higher-quality online experience; global TD learning and local preference ranking progressively improve the critic; and relative-advantage optimization exploits the calibrated critic without aggressively pursuing absolute $Q$ values. The reference policy limits the extent to which critic errors can alter previously acquired task competence.

ACoB-Stream and the experience–policy bottleneck

The systems contribution is structured around the lifecycles of two state classes: experience-context state and policy state. The architecture is designed for settings in which the frozen multimodal prefix is computationally expensive but invariant during Stage II.

The first mechanism is cross-update context reuse. During actor inference, the system stores compacted prefix key–value caches for each observation. Replay and correction records retain identifiers rather than duplicating the cached contexts. Because the prefix is frozen, the learner can train the action expert without recomputing the multimodal prefix for every sampled transition. The computational cost of prefix formation consequently changes from once per sampled item to once per new observation.

The second mechanism is deduplicated disk-backed persistence. Contexts are stored once, while replay and correction buffers reference them by identifier. Sliding-window sampling constrains the active working set to the available operating-system page cache, improving locality as the replay history grows. The design retains the complete disk-backed history for recovery while avoiding the random-access behavior of uniformly sampling a growing replay store.

The third mechanism is objective-aligned access. Critic updates require successor contexts for bootstrap targets, whereas action-expert updates require both current and successor contexts. ACoB-Stream therefore avoids retrieving context tensors that are irrelevant to the current objective. CPU workers sample transitions and assemble batches while GPU workers consume prefetched batches, reducing learner idle time.

The fourth mechanism is trainable-subspace synchronization. The invariant VLA state remains resident in both actor and learner processes. Only the updated action-expert state is transmitted and merged into the actor. The actor continues using the previous valid policy during preparation and atomically switches to the new state after loading. This reduces synchronization bandwidth and policy-version lag without requiring full-model serialization.

Figure 1: ACoB-Stream manages experience-context and policy-state lifecycles through reuse, persistence, objective-aligned access, and trainable-subspace synchronization.

The architecture therefore targets the complete closed loop rather than an isolated kernel. Its objective is not merely faster gradient computation, but faster conversion of physical experience into a deployed policy update.

Experimental design and physical task suite

The evaluation covers nine chemistry-manipulation tasks in four categories: contact-rich manipulation, contact-light manipulation, contact-free manipulation, and bimanual coordination. The suite includes vial and cuvette transfer, tube-rack loading, rubber-stopper insertion, alcohol-lamp extinguishing, pipette-tip attachment, bulb-dropper transfer, pipette transfer and ejection, and tube brushing.

These tasks impose different sources of difficulty. Transparent objects and narrow openings stress visual localization. Insertion and attachment tasks require millimeter- or submillimeter-scale alignment and appropriate force control. Long sequences expose accumulated action errors. The bimanual tube-brushing task tests coordination between independently controlled arms. Object poses and initial robot configurations are randomized across evaluation trials.

Experiments use four embodiments: a UR5e with a parallel gripper, a UR5e with a dexterous hand, a dual-UR5e platform, and a Franka Research 3 with a parallel gripper. Single-arm systems use one wrist camera and one external camera; the dual-arm system uses two wrist cameras and one external camera.

Figure 2: The four evaluated hardware configurations, including single-arm, dexterous-hand, dual-arm, and Franka platforms.

The authors also construct a six-degree-of-freedom active isomorphic master for UR5e teleoperation. Its graded actuation assigns higher torque to proximal joints and lower inertia to distal joints, while an independent shaft and dual-bearing support decouple structural loads from servo actuation. For tasks requiring fine alignment but limited orientation variation, the system uses a Cartesian keyboard interface with coarse and fine increments. Both interfaces are converted into a common step-wise delta task-space representation and recorded at 15 Hz.

The baselines include HIL-SERL, ConRFT, Robo-Dopamine, and demonstration-only fine-tuning of $\pi_0$ and $\pi_{0.5}$. The comparison is consequential because the methods differ in model scale, policy parameterization, reward construction, and initialization. The reported results should therefore be interpreted as an end-to-end comparison under the paper’s task and hardware protocols, rather than as an isolated algorithmic comparison with all factors perfectly controlled.

Main real-world results

Across all nine tasks, VLA-Precision achieves a mean success rate of 98.3% after 45.8 minutes of online training per task, completing successful episodes in 27.6 seconds on average. It succeeds in 177 of 180 held-out trials, reaches 100% success on seven tasks, and maintains at least 90% success on every task.

Method	Mean success	Mean episode time
HIL-SERL	2.8%	50.4 s
ConRFT	7.8%	45.3 s
Robo-Dopamine	10.0%	44.5 s
$\pi_0$	59.4%	32.0 s
$\pi_{0.5}$	67.8%	30.0 s
VLA-Precision	98.3%	27.6 s

Relative to $\pi_{0.5}$0, the framework improves mean success by 30.5 percentage points and reduces mean episode time by 8.7%. Relative to $\pi_{0.5}$1, it improves success by 38.9 percentage points and speed by 15.9%. The worst-task success rate is 90%, compared with 10% for $\pi_{0.5}$2 and 5% for $\pi_{0.5}$3. The paper reports an 88.3% relative improvement in mean success and a 61.2% relative speed improvement over Robo-Dopamine.

Figure 3: Held-out success rates and completion times on representative tasks, with Wilson confidence intervals for success rates.

The result is strongest on tasks for which demonstration-only policies retain partial competence but fail under pose variation or long-horizon precision demands. For example, VLA-Precision reaches 100% on tube-rack loading, 2 mL vial transfer, rubber-stopper insertion, alcohol-lamp extinguishing, pipette-tip attachment, bulb-dropper transfer, and pipette transfer and ejection. It reaches 90% on cuvette transfer and 95% on tube brushing.

The comparison with the two large VLA baselines indicates that pretrained general-purpose priors are valuable but insufficient. Full fine-tuning of $\pi_{0.5}$4 achieves 67.8% mean success, showing substantial retained competence, but remains limited by the demonstration distribution. ACoB’s improvement to 98.3% supports the paper’s claim that real-world interaction can overcome a demonstration-limited ceiling when policy updates are sufficiently conservative and value signals are calibrated.

The training dynamics provide a complementary view. On five representative tasks, the final autonomous success rate averages 98.0%, while the effective success rate averages 99.0%. The intervention rate averages only 0.024%. Compared with Robo-Dopamine, the authors report increases of 92.0 and 49.5 percentage points in autonomous and effective success, respectively, together with a 66.9% reduction in interventions.

Figure 4: Online training dynamics showing autonomous and effective success, cumulative interventions, and intervention rate.

The task-level results also expose the specific failure pattern of the baselines. On multistep vial transfer, ConRFT and Robo-Dopamine reduce intervention rates relative to HIL-SERL but remain at only 5% autonomous success. On long-horizon pipette transfer and ejection, all three real-world RL baselines remain at 0% autonomous success. On bimanual tube brushing, the baselines likewise achieve 0% autonomous success despite offline task-specific training. These results are consistent with the paper’s diagnosis that BC regularization alone can preserve an initial behavior while absolute-$\pi_{0.5}$5 optimization still accumulates action errors and induces policy drift.

The paper’s offline evaluation on five representative tasks reports 99 successes out of 100 trials for VLA-Precision, compared with 61 for $\pi_{0.5}$6 and 50 for $\pi_{0.5}$7. Mean completion time is 26.65 seconds for VLA-Precision, versus 29.32 and 32.05 seconds for $\pi_{0.5}$8 and $\pi_{0.5}$9, respectively. Because VLA-Precision uses fewer Stage-I demonstrations and fewer Stage-I optimization steps than the demonstration-only baselines before online RL, the result also suggests that online data compensate for a comparatively smaller initial supervised dataset. However, this comparison combines different training budgets and procedures, so it does not isolate the marginal effect of online RL independently of all initialization choices.

Throughput and systems evaluation

The ACoB-Stream evaluation compares five systems: a no-KV-cache variant, full-history random disk sampling, naive context access, a dynamic CPU-RAM alternative, and the complete disk-based ACoB-Stream system.

The complete system performs 15,666 critic-to-actor cycles in 150 minutes, corresponding to 1.7407 cycles per second and 0.574 seconds per cycle. Relative to the no-KV-cache system, it provides 10.95 times higher throughput and reduces mean latency by 90.9%. Relative to full-history random sampling, it provides 1.53 times higher throughput and 34.5% lower mean latency. Relative to naive context access, it provides 1.80 times higher throughput and 44.4% lower mean latency.

Figure 5: ACoB-Stream improves critic-to-actor throughput and latency through context reuse, locality-aware persistence, and objective-aligned retrieval.

The random-sampling ablation begins degrading after approximately 43 minutes, when the replay working set exceeds Linux page-cache capacity. Its throughput falls from 1.56 to 0.45 cycles per second, whereas ACoB-Stream reaches 1.74 cycles per second in the final five-minute interval. This observation supports the claim that disk persistence alone is insufficient: the sampling policy must preserve access locality as the experience store grows.

The dynamic CPU-RAM alternative remains competitive, but the full disk-based system achieves 1.09 times its throughput and 8.0% lower latency while avoiding the additional coordination required for dynamic active-window sizing, FIFO eviction, and disk persistence. The magnitude of the no-cache comparison indicates that frozen-prefix reuse is the dominant systems optimization in this implementation, while the other mechanisms preserve performance as the replay history expands.

Ablation of ACoB mechanisms

The algorithmic ablations provide unusually large separations among the components. Full ACoB reaches 96.25% final autonomous success with a 0.03% intervention rate averaged across four tasks. Removing critic preference reduces autonomous success to 26.25%; removing actor behavior cloning reduces it to 20.00%; and replacing relative advantage with direct sampled-action $Q(\omega,a)=V(\omega)+A(\omega,a)$0 maximization reduces it to 8.75%.

Variant	Autonomous success	Intervention rate	Cumulative interventions
Full ACoB	96.25%	0.03%	10,080
Without critic preference	26.25%	10.57%	16,580
Without actor BC	20.00%	9.24%	19,581
Without relative advantage	8.75%	24.93%	28,806

Figure 6: Removing critic preference, intervention-guided behavior cloning, or relative-advantage improvement substantially degrades autonomous learning.

The critic-preference ablation supports the claim that executed-action TD learning does not adequately assign credit when interventions replace policy proposals. Without preference ranking, the critic can observe the successful correction but lacks a state-matched negative signal for the original proposal. The 70.00-point success decrease indicates that this distinction is operationally important in the evaluated tasks.

The actor-BC ablation shows that relying on indirect critic guidance is insufficient during early online learning. Without direct absorption of successful and corrected actions, the policy improves more slowly, generates poorer experience, and requires substantially more intervention. The result empirically supports the paper’s cross-timescale argument: behavioral cloning is not merely a stabilizing constraint but an active mechanism for improving the data used by value learning.

The largest degradation occurs when relative advantage is replaced by absolute $Q(\omega,a)=V(\omega)+A(\omega,a)$1 maximization. Autonomous success drops by 87.50 percentage points and intervention rate rises to 24.93%. This result is consistent with the paper’s central diagnosis of optimistic value errors and scale sensitivity. It also provides the strongest evidence that the benefit of ACoB does not arise solely from human corrections or reference regularization; the form of the policy-improvement signal is critical.

Action horizon and error accumulation

The paper evaluates step-wise delta actions at different executed horizons on three precision-insertion tasks. Increasing the executed horizon from $Q(\omega,a)=V(\omega)+A(\omega,a)$2 to $Q(\omega,a)=V(\omega)+A(\omega,a)$3 reduces mean autonomous success from 58.3% to 6.7% and increases relative action-prediction RMSE by 29.3%.

Figure 7: Longer executed action horizons increase action-prediction error, particularly during interaction-critical phases, and sharply reduce autonomous success.

This result motivates the use of $Q(\omega,a)=V(\omega)+A(\omega,a)$4 in the main experiments. Step-wise deltas permit frequent closed-loop correction and keep individual targets bounded, but their errors accumulate recursively across the action horizon. The paper observes that marginal RMSE increases remain positive as the horizon is extended and are larger during interaction-critical phases than during normal motion. The implication is specific and important: the demonstrated success of VLA-Precision depends partly on a short execution horizon, and the same action representation cannot be assumed to support substantially longer-horizon bimanual manipulation.

Limitations and open questions

The evaluation is broad in task count and embodiment coverage, but it remains task-specific. Only one bimanual task is included, and the paper does not establish performance on longer bimanual sequences with multiple precision-critical stages. The action-horizon analysis itself shows that step-wise deltas become unreliable as the executed horizon grows, with success falling from 58.3% to 6.7% between horizons of three and six. Whether chunk-wise deltas, hierarchical action representations, or another state parameterization can preserve ACoB’s stability over longer sequences remains open.

The experiments also rely on substantial human infrastructure: task demonstrations, intervention data, custom teleoperation hardware, reward and success labeling, and task-specific critic ensembles. The paper does not quantify the total human labor required per task or compare intervention quality across operators. Its claim of efficient online training is therefore primarily a claim about robot interaction time and learner throughput, not a complete accounting of engineering and annotation cost.

The baseline comparison is informative but not fully controlled across model families and optimization protocols. The Octo-based methods, $Q(\omega,a)=V(\omega)+A(\omega,a)$5, and $Q(\omega,a)=V(\omega)+A(\omega,a)$6 differ in architecture, pretraining, action representation, and task initialization. In addition, VLA-Precision uses fewer Stage-I demonstrations than the demonstration-only baselines but introduces online interventions and a custom systems stack. The reported superiority establishes strong end-to-end performance on the selected suite, while leaving the contribution of each model and data advantage less precisely identified.

Finally, ACoB uses task-specific critics and task-specific online training. The paper proposes multi-task extension as an open direction but does not evaluate transfer, interference, catastrophic forgetting, or cross-task critic calibration. It is therefore not yet established whether the asymmetric co-bootstrapping mechanism remains stable when experience from multiple tasks shares an action expert and a replay system.

Conclusion

VLA-Precision presents a coordinated algorithmic and systems solution to real-world online RL for large VLAs. ACoB combines intervention-driven behavioral learning, global return propagation, state-matched critic preference ranking, pessimistic relative advantages, and reference regularization. ACoB-Stream makes this loop computationally viable through frozen-prefix reuse, deduplicated persistence, objective-aligned context access, and trainable-subspace synchronization.

On nine high-precision chemistry tasks and four embodiments, the framework reports 98.3% mean success after 45.8 minutes per task, 27.6-second successful episodes, and 10.95-fold higher learner throughput than a no-cache implementation. The ablations indicate that all three principal ACoB mechanisms are necessary for the reported autonomous performance, while the horizon study identifies a concrete boundary imposed by step-wise delta actions. The paper’s principal unresolved question is whether these gains extend to multi-task and substantially longer-horizon bimanual online RL without sacrificing the value calibration and behavioral stability on which the current results depend.