VLA-Precision: Framework Overview

• The VLA-Precision framework applies reinforcement-learning methods to improve the precision, repeatability, and efficiency of pretrained vision-language-action models in real-world robotic manipulation, achieving a 98.3% mean success rate across nine high-precision chemistry tasks.
• ACoB, a core component, integrates behavioral learning and calibrated value estimation for stable policy improvement, while ACoB-Stream reduces computational costs through efficient experience policy architecture.
• It significantly enhances efficiency, reaching up to 10.9X improvement in training throughput and computational efficiency

VLA-Precision is an online reinforcement-learning framework for improving the precision, repeatability, and efficiency of pretrained vision-language-action models in real-world robotic manipulation. It combines Asymmetric Co-Bootstrapping (ACoB), which integrates intervention-guided behavioral learning with progressively calibrated value estimation, and ACoB-Stream, a closed-loop experience–policy architecture for reducing the computational cost of large-VLA training. The framework is evaluated on nine high-precision chemistry tasks across four robot embodiments and reports a mean success rate of $98.3\%$, an average successful-episode time of $27.6$ s, and up to $10.9\times$ improvement in training throughput and computational efficiency (Su et al., 3 Sep 2026).

1. Problem setting and objectives

VLA-Precision addresses a two-stage adaptation problem. In Stage I, a pretrained flow-based VLA, denoted $\pi_{0.5}$, is fully fine-tuned on task demonstrations. This produces a competent task-specific policy with visual-language priors and an action expert adapted to the robot and task. In Stage II, the robot performs real-world online RL, optionally receives human corrections, and updates the action expert using physical outcomes.

The physical task is modeled as an MDP,

$$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$

where the language-conditioned state is $s_t=(o_t,q_t,\ell)$, with visual observation $o_t$, robot state $q_t$, and language instruction $\ell$. Each action is an $H$-step action chunk,

$27.6$0

with chunk reward

$27.6$1

The online objective is

$27.6$2

where

$27.6$3

The parameters are partitioned into frozen multimodal-prefix parameters $27.6$4, frozen Stage-I action-expert parameters $27.6$5, and trainable LoRA parameters $27.6$6. The action expert produces stochastic action chunks according to

$27.6$7

with $27.6$8.

The central difficulty is that precision manipulation is highly sensitive to small policy changes. Sparse or delayed physical rewards can produce inaccurate value estimates, while direct maximization of an overestimated critic can cause policy drift away from the capable Stage-I model. Human interventions introduce an additional credit-assignment problem: the successful outcome may be caused by the corrective action rather than by the VLA’s original proposal, but conventional temporal-difference learning does not necessarily penalize the overwritten proposal.

Behavior cloning alone has the opposite limitation. It can reproduce demonstrations rapidly but is vulnerable to distribution shift and compounding errors, and it cannot systematically improve beyond the demonstrator. VLA-Precision therefore combines behavioral learning, return propagation, local action comparisons, and reference-policy regularization.

2. Asymmetric Co-Bootstrapping

ACoB is a cross-timescale learning algorithm. Early policy improvement is dominated by intervention-guided behavioral learning, which rapidly converts human corrections into better autonomous behavior. As experience accumulates, global return propagation and local preference ranking calibrate the critics. The resulting relative advantages are then used for conservative policy improvement.

The actor–learner loop maintains three experience structures:

• Replay buffer $27.6$9: all transitions.
• Correction buffer $10.9\times$0: effective human interventions.
• Context buffer $10.9\times$1: stored frozen-prefix contexts.

A transition is represented as

$10.9\times$2

where $10.9\times$3 is the terminal indicator and

$10.9\times$4

The VLA proposal and executed action are explicitly distinguished:

$10.9\times$5

and

$10.9\times$6

An intervention is effective when $10.9\times$7 and the executed action differs from the proposal by more than a tolerance:

$10.9\times$8

The effective-intervention label is used both for behavioral training and for critic preference supervision.

Progressive value calibration

ACoB uses an ensemble of $10.9\times$9 critics. Each critic decomposes the action value into a state value and an action advantage:

$\pi_{0.5}$0

where

$\pi_{0.5}$1

The raw action is normalized into the VLA’s OpenPI representation and then mapped to the critic representation:

$\pi_{0.5}$2

For global return propagation, the learner samples a bootstrap action at the successor state and constructs the pessimistic target

$\pi_{0.5}$3

The critic TD loss is

$\pi_{0.5}$4

This component propagates long-horizon consequences to earlier executed actions.

TD learning alone does not compare the executed corrective action with the VLA proposal at the same state. ACoB therefore adds local preference ranking for effective interventions:

$\pi_{0.5}$5

The ranking loss is

$\pi_{0.5}$6

where $\pi_{0.5}$7 and $\pi_{0.5}$8 is the ranking margin. The ranking is applied to $\pi_{0.5}$9, rather than directly to $$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$0, because the state value $$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$1 is shared by both actions. The complete critic objective is

$$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$2

with reported shared setting $$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$3.

Relative-advantage policy improvement

Rather than directly maximizing an absolute critic value, ACoB compares the current policy against a conservative baseline. The current policy and frozen Stage-I reference policy use the same noise sample:

$$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$4

$$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$5

where $$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$6 is the frozen Stage-I LoRA state.

For $$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$7,

$$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$8

The critic-specific baseline is

$$\mathcal M_\ell=(\mathcal S,\mathcal A,P,r,\rho_0,\bar\gamma),$$9

Without an effective correction, the baseline is the frozen-reference advantage. With an effective correction, it becomes effectively the larger of the reference and proposal advantages. The pessimistic relative advantage is

$s_t=(o_t,q_t,\ell)$0

The actor loss is a smooth margin objective:

$s_t=(o_t,q_t,\ell)$1

with $s_t=(o_t,q_t,\ell)$2. The ensemble minimum requires the proposed policy action to outperform the baseline according to every critic.

Intervention-guided behavioral learning

ACoB uses flow-matching behavioral learning as its fast timescale. For a demonstrated executed action chunk,

$s_t=(o_t,q_t,\ell)$3

and

$s_t=(o_t,q_t,\ell)$4

The behavioral mask is

$s_t=(o_t,q_t,\ell)$5

where $s_t=(o_t,q_t,\ell)$6 indicates episode success. Consequently, behavioral learning uses all chunks from successful episodes and only effective corrections from failed episodes:

$s_t=(o_t,q_t,\ell)$7

This mechanism improves the policy before the value estimates become reliable, reducing the number of future interventions and improving the quality of subsequent autonomous data.

Frozen-reference regularization

The Stage-I policy is retained as a frozen reference. ACoB regularizes body-action coordinates toward that reference:

$s_t=(o_t,q_t,\ell)$8

The complete action-expert objective is

$s_t=(o_t,q_t,\ell)$9

with

$o_t$0

The reference penalty is not expressed as an explicit KL constraint, but serves a related trust-region function by discouraging broad distributional changes caused by transient critic errors.

3. ACoB-Stream systems architecture

ACoB-Stream addresses the computational bottlenecks of online RL with large VLAs. Standard actor–learner systems repeatedly recompute the frozen multimodal prefix, transfer large KV states, perform random access into growing replay histories, and synchronize full model states between learner and robot actor.

ACoB-Stream is organized around two principles:

1. Invariant-state decoupling: frozen-prefix states are reused because the multimodal prefix remains unchanged during Stage II.
2. On-demand streaming: contexts are retrieved according to the current optimization objective rather than loaded indiscriminately.

Context formation and persistence

During actor inference, the frozen multimodal-prefix KV context is stored rather than recomputed during every learner update. At episode commitment:

• all transitions enter $o_t$1;
• effective corrections enter $o_t$2;
• compacted prefix contexts enter $o_t$3;
• transitions store identifiers pointing to current and successor contexts in $o_t$4.

Inactive observation positions are removed before persistence. The context buffer is disk-backed and deduplicated, while replay records store identifiers rather than duplicated KV tensors.

A sliding-window sampler maintains the active working set near Linux page-cache capacity. This preserves locality while retaining the full history on disk. The method contrasts with uniform sampling over the entire disk history, which causes increasing page-cache misses and disk-to-CPU transfers as replay grows.

Objective-aligned retrieval

Different optimization objectives require different contexts:

• critic TD updates require successor contexts for bootstrap targets;
• actor updates require current contexts and, where necessary, successor contexts;
• preference and correction updates require contexts associated with both proposal and executed actions.

CPU workers sample transitions, retrieve and batch the relevant contexts, and prepare host-side data. Device workers transfer complete batches to the GPU while prefetching the next batch.

Policy-state synchronization

The frozen VLA state remains resident in both actor and learner. The learner publishes only the trainable action-expert state rather than serializing the entire VLA. The actor merges the updated trainable state into the local VLA, performs device placement, and atomically swaps the active state under a lock. Rollouts can continue using the previous valid version while the new version is prepared.

Throughput results

A CTA cycle consists of one critic-only update and one joint critic–actor update. The full Disk ACoB-Stream system achieves

$o_t$5

corresponding to

$o_t$6

Relative to system ablations, ACoB-Stream provides:

• $o_t$7 throughput and $o_t$8 lower mean latency than the configuration without KV caching;
• $o_t$9 throughput and $q_t$0 lower latency than Disk Full Random;
• $q_t$1 throughput and $q_t$2 lower latency than Naive Context Access;
• $q_t$3 throughput and $q_t$4 lower latency than Dynamic CPU RAM.

The maximum reported improvement is up to $q_t$5 in throughput and computational efficiency.

4. Experimental platform and task suite

VLA-Precision is evaluated on four robot embodiments:

1. UR5e with a PGI-140-80 parallel gripper;
2. UR5e with a LinkerHand L20 dexterous hand;
3. dual UR5e arms, each with a PGI-140-80 gripper;
4. Franka Research 3 with a PGI-140-80 gripper.

Single-arm systems use one wrist camera and one external camera. The dual-arm system uses two wrist cameras and one external camera. Training uses four NVIDIA A800 GPUs.

The benchmark contains nine chemistry tasks in four categories.

Contact-rich tasks include Tube Rack Loading, 2 mL Vial Transfer, and Cuvette Transfer. These require controlled interaction with laboratory objects and containers.

Contact-light tasks include Rubber Stopper Insertion, Alcohol Lamp Extinguishing, and Pipette Tip Attachment.

Contact-free tasks include Bulb Dropper Transfer and Pipette Transfer and Ejection.

Bimanual manipulation is represented by Tube Brushing.

Object poses and initial arm poses are varied, with initial reset perturbations of $q_t$6 cm in $q_t$7. Data are recorded at a nominal $q_t$8 Hz. The control representation uses step-wise delta task-space actions,

$q_t$9

and

$\ell$0

Stage-I training uses $\ell$1–$\ell$2 demonstrations per task and $\ell$3–$\ell$4 training steps for VLA-Precision. The baseline $\ell$5 and $\ell$6 models use $\ell$7–$\ell$8 demonstrations and $\ell$9–$H$0 supervised fine-tuning steps.

Human interventions are provided either through a six-degree-of-freedom active isomorphic master or a Cartesian keyboard interface. The former is used for longer sequences and compliant interaction; the latter supplies coarse and fine translational or rotational corrections suitable for millimeter and submillimeter alignment.

The principal comparison methods are HIL-SERL, ConRFT, Robo-Dopamine, $H$1, and $H$2. Evaluation includes task success, autonomous success, effective success, intervention count and rate, successful-trial episode time, training time, CTA throughput, and critic-to-actor latency.

5. Quantitative performance

Across the complete nine-task suite, VLA-Precision reports:

Method	Mean success	Mean episode time
HIL-SERL	$H$3	$H$4 s
ConRFT	$H$5	$H$6 s
Robo-Dopamine	$H$7	$H$8 s
$H$9	$27.6$00	$27.6$01 s
$27.6$02	$27.6$03	$27.6$04 s
VLA-Precision	$27.6$05	$27.6$06 s

VLA-Precision succeeds in $27.6$07 held-out trials, reaches $27.6$08 on seven tasks, and reaches at least $27.6$09 on every task. Relative to $27.6$10, mean success improves by $27.6$11 percentage points and average execution time decreases by $27.6$12. Relative to $27.6$13, the success improvement is $27.6$14 percentage points and the speed improvement is $27.6$15.

On five representative tasks, the final averages are:

• autonomous success: $27.6$16;
• effective success: $27.6$17;
• cumulative interventions: $27.6$18;
• intervention rate: $27.6$19.

In an offline evaluation over five tasks, VLA-Precision achieves $27.6$20 successes, compared with $27.6$21 for $27.6$22 and $27.6$23 for $27.6$24. Mean successful-trial time is $27.6$25 s for VLA-Precision, $27.6$26 s for $27.6$27, and $27.6$28 s for $27.6$29.

Algorithmic ablations

Across four representative tasks, the full ACoB system achieves $27.6$30 final autonomous success, a $27.6$31 intervention rate, and $27.6$32 interventions.

Variant	Autonomous success	Intervention rate	Interventions
Full ACoB	$27.6$33	$27.6$34	$27.6$35
Without critic preference	$27.6$36	$27.6$37	$27.6$38
Without actor BC	$27.6$39	$27.6$40	$27.6$41
Without relative advantage	$27.6$42	$27.6$43	$27.6$44

Removing critic preference causes the critic to lose explicit supervision distinguishing successful corrections from inferior proposals. Removing actor behavioral learning prevents rapid incorporation of interventions and forces the policy to rely on slower critic feedback. Replacing relative-advantage improvement with direct sampled-action $27.6$45-maximization is the most damaging ablation, consistent with increased exposure to optimistic value errors and policy drift.

Generalization across tasks and embodiments

The reported results span contact-rich, contact-light, contact-free, and bimanual manipulation, as well as four embodiments. The framework also evaluates varying object poses and initial arm configurations, transparent, fragile, deformable, and narrow-tolerance objects, and long-horizon chemistry procedures.

The results are task-specific: each task is trained separately. Multi-task real-world RL is identified as future work.

6. Precision, stability, and limitations

VLA-Precision defines precision operationally through repeatable task success, effective contact behavior, intervention reduction, and safe execution time rather than through a single geometric error metric. Its main mechanisms address different aspects of this objective.

Intervention-guided behavioral learning rapidly transfers corrective actions into the policy. This reduces repeated failures and improves the distribution of subsequent autonomous experience.

Local preference ranking gives explicit credit to the corrective action over the original proposal at the same state. This is especially important when a human intervention rescues a rollout.

Global return propagation assigns long-horizon consequences to earlier executed actions through TD bootstrapping.

Relative advantages avoid reliance on the absolute scale of an imperfect critic. The current policy must improve relative to a frozen reference and, when relevant, relative to the original proposal.

Frozen-reference regularization preserves Stage-I manipulation priors and suppresses uncontrolled policy drift.

ACoB-Stream increases the rate at which physical experience can influence the policy by reusing invariant multimodal-prefix states, reducing KV recomputation, limiting context retrieval to the current objective, and transmitting only trainable action-expert parameters.

Several limitations remain. The experiments are task-specific and do not establish multi-task real-world online RL. Long-horizon step-wise delta actions can accumulate error. Increasing the executed action horizon $27.6$46 from $27.6$47 to $27.6$48 reduces average autonomous success from $27.6$49 to $27.6$50 on three precision-insertion tasks and increases relative action-prediction RMSE by $27.6$51. The paper therefore identifies chunk-wise action representations as a possible requirement for longer bimanual sequences.

The system also depends on human intervention quality, reward design, critic calibration, and safe exploration. It uses bounded delta actions, impedance control, a frozen reference policy, ensemble-based conservative advantages, and human takeover interfaces, but does not provide a formal constrained-RL or reachability-based safety guarantee.

The paper does not independently tabulate the contribution of reference regularization, nor does it provide all critic architecture specifications, every task-specific reward definition, all values of $27.6$52, $27.6$53, and $27.6$54, or complete optimizer settings. Its quantitative evidence is strongest for the combined ACoB and ACoB-Stream system within the evaluated chemistry suite.

VLA-Precision’s principal significance is the integration of algorithmic and systems-level design for real-world VLA adaptation. ACoB stabilizes online policy improvement by combining fast intervention learning with progressively calibrated relative value estimates. ACoB-Stream makes this process computationally viable by treating frozen multimodal context as reusable state rather than repeatedly recomputed model input. Together, these mechanisms produce high reported success and low intervention rates while retaining the pretrained VLA’s manipulation competence. The framework therefore characterizes precision not as a property of the action expert alone, but as the result of controlled policy improvement, explicit proposal–correction comparison, reference preservation, and efficient closed-loop interaction.