Variational Automation (VA) in Robotics

• Variational Automation is a regime between fixed automation and generalist robotics, handling tasks with substantial variations in object geometry and pose while ensuring persistent, reliable execution.
• The Graph-as-Policy representation constructs explicit, modular computation graphs from robotic skills, enhancing policy interpretability and recovery through structured node and edge management.
• Empirical benchmarks show that VA systems, utilizing internal simulation and iterative self-learning, outperform model-free policies by maintaining high success rates and sustained throughput.

Variational Automation (VA) denotes a class of robot tasks with larger variations in object geometry and pose than fixed automation, but with the persistent execution and reliability requirements of commercial and industrial automation. The term is introduced to describe a regime between “fixed automation” and “generalist robotics,” with representative tasks including grocery fulfillment, cable insertion, and crate washing. In this setting, the central problem is not only task completion, but also robust execution under structured variation, interpretable policy construction, and sustained throughput over repeated deployments (Chen et al., 6 Jul 2026).

1. Definition and task regime

VA is defined as a class of tasks that have larger variations in object geometry and pose than fixed automation. The defining contrast is with two neighboring regimes: fixed automation, characterized as repetitive tasks with minimal variation, and generalist robotics, characterized as highly unstructured open-world tasks. VA therefore refers to structured, persistent automation with significant variations in object geometry and pose, rather than to open-ended autonomy in unconstrained environments (Chen et al., 6 Jul 2026).

The motivation for isolating this regime is operational rather than purely taxonomic. Recent work asks whether agentic coding systems can combine interpretable robot programming with the open-world adaptability of model-free policies, because model-free policies often struggle to close the reliability gap for VA tasks, which must be executed persistently and reliably in commercial and industrial applications. The term thus emphasizes variation under deployment constraints: the task distribution is broader than in fixed automation, but the acceptable failure rate remains far tighter than in many research benchmarks (Chen et al., 6 Jul 2026).

Examples given for VA include grocery fulfillment, cable insertion, and crate washing. These are not presented as isolated demonstrations; they are used to motivate a broader class of automation problems in which object geometry, pose, and task instance vary substantially, while throughput and repeatability remain primary requirements (Chen et al., 6 Jul 2026).

2. Benchmark suite and evaluation criteria

A concrete instantiation of VA is provided through 8 new open VA task benchmarks, with 4 in simulation and 4 in the real world. These are intended to capture larger object/pose variation than standard fixed-automation tasks such as LIBERO, while preserving the practical emphasis on persistent task execution (Chen et al., 6 Jul 2026).

Benchmark family	Simulation	Real-world
Fulfill Grocery Orders	I-a	I-b
Pack Grocery Items	II-a	II-b
Make Popcorn	III-a	III-b
Wash Crates (Bimanual)	V	—
Insert USB-C Cables	—	IV

The evaluation emphasizes success rate and sustained throughput. This is consequential because VA is framed as a deployment-oriented regime: success on isolated episodes is insufficient if cycle time degrades or if the policy fails under routine geometric and pose perturbations. Reported task variants include object, pose, and goal variations, and some comparisons also examine increased language variation (Chen et al., 6 Jul 2026).

The benchmark design also supports iterative rehearsal. Task instances are sampled from a belief space $\mathcal{B}$, described as a distribution over object poses and environment state, so that candidate policies can be stress-tested over varied scenes rather than only over fixed demonstrations (Chen et al., 6 Jul 2026).

3. Graph-as-Policy representation

The principal policy formalism proposed for VA is Graph-as-Policy (GaP), a multi-agent coding harness that generates directed computation graphs with perception, planning, and control nodes from a Modular Open Robot Skill Library (MORSL). In this representation, policies are not monolithic scripts or end-to-end latent policies; they are explicit directed graphs $\mathcal{G} = (V, E)$ whose nodes are robot skills and whose edges encode dependencies (Chen et al., 6 Jul 2026).

The paper formalizes execution as

$$\pi(a \mid \mathbf{x}, \mathcal{T}) = \text{execution of computation graph } \mathcal{G} \text{ over inputs}~\mathbf{x}, \mathcal{T}.$$

The associated performance criterion combines success and throughput:

$$\mathcal{J} = w_s \cdot \mathbb{I}(\text{success}) + w_t \cdot \Phi,$$

where $\Phi$ is the throughput (Chen et al., 6 Jul 2026).

Within this formalism, nodes represent atomic robot skills with strict input/output types. The core node types are perception nodes, planning nodes, control nodes, logic/control-flow nodes, and composite nodes. Perception nodes segment, detect, or localize objects and subparts via VLMs or SAM and may output masks, oriented bounding boxes, or point clouds. Planning nodes compute grasp poses, motion plans, or manipulation waypoints through tools such as cuRobo, M2T2, and GraspGen. Control nodes execute robot, joint, or gripper motions, follow trajectories, or call learned policies. Logic/control-flow nodes route execution conditionally, while composite nodes package multi-step procedures into reusable subgraphs (Chen et al., 6 Jul 2026).

Edges are either data edges or control edges. Data edges transmit outputs from one node to inputs of another; control edges implement conditional routing based on node outputs such as success or failure. This distinction makes the policy explicitly interpretable and also supports recovery logic, completion checks, and modular reuse (Chen et al., 6 Jul 2026).

MORSL functions as the declarative skill library underlying GaP. It is described as a composable, extensible library of 51+ skills spanning perception, model-free grasping, motion planning, geometry, vision, control, and verification. Each skill specifies type-checked inputs/outputs, semantic parameters, and pre-conditions. Skills can call ROS nodes, custom scripts, or gRPC interfaces. The graph synthesis process is handled by an Orchestration Agent, which decomposes the task into sub-tasks, and Skill Agents, which synthesize local subgraphs that are subsequently aggregated, type-checked, and validated before deployment (Chen et al., 6 Jul 2026).

4. Internal simulation and self-learning

GaP is not limited to one-shot graph synthesis. It generates an internal simulation environment to rehearse task instances with different graphs in parallel and to iteratively refine graph structure and parameters to improve success rates and throughput. The simulator is described as physics-based, with NVIDIA Isaac given as an example (Chen et al., 6 Jul 2026).

The self-learning loop operates over the belief space $\mathcal{B}$. At each iteration, the system samples task instances, rehearses execution of the current graph in parallel, analyzes failures, and performs graph revision through LLM-based agents. The pseudocode names this sequence through functions such as $\mathrm{Graph\_Init}$, $\mathrm{Build\_Scene}$, $\mathrm{Rehearsal}$, $\mathrm{Analyze\_Failure}$, and $\mathcal{G} = (V, E)$0 (Chen et al., 6 Jul 2026).

Failure analysis is granular. The simulation logs node-level successes and failures, including examples such as “grasp failed, pan not grasped” and “placement offset insufficient coverage of burner footprint.” Postcondition checkpoints, trajectory and pose comparison, and object contact checks are used to determine where the graph should be revised. Revisions can include node replacement, rewiring, or parameter modification (Chen et al., 6 Jul 2026).

This design makes the policy both interpretable and self-improving. Because the graph is explicit, corrective changes can target specific perception, planning, or control failures rather than retraining an opaque policy end to end. Because rollouts are parallelized, optimization is framed around repeated rehearsal across varied task instances, rather than around manual debugging of individual failures (Chen et al., 6 Jul 2026).

5. Empirical performance on VA tasks

Evaluation on the 8 VA benchmarks suggests that GaP can achieve success rates that significantly outperform baselines. The reported comparisons include model-free policies, TAMP systems, and prior agentic coding approaches, and the metrics emphasize both success rate and sustained throughput (Chen et al., 6 Jul 2026).

On grocery fulfillment and packing tasks, GaP achieves success rates of 0.93–0.99 on diverse variants with object, pose, and goal variation. Under increased pose and object variation, model-free VLAs such as $\mathcal{G} = (V, E)$1 and MolmoAct2 drop to 0.10–0.20. On real-robot grocery order fulfillment (I-b), GaP achieves 25/25 (100%) success, while TipTop obtains 8/25; the summary also reports baseline ranges of 0.01–0.24 for CaP-X, 0.10–0.43 for VLA, and 0.22–0.46 for TAMP/TipTop on the compared tasks (Chen et al., 6 Jul 2026).

On the Wash Crates benchmark, GaP matches an expert hand-tuned policy closely: 0.95 success with 179s per cycle ($\mathcal{G} = (V, E)$218/hr), versus 0.99 success and 176s for the human policy. In USB-C Insertion (IV), GaP attains 0.93 (121/130) across multiple goal and pose prompts and remains robust to cable orientation and port pose changes (Chen et al., 6 Jul 2026).

The most explicit demonstration of self-learning is reported for Make Popcorn. The initial graph achieves a 33% success rate, with failures in pan grasp and placement. After 10 self-learning iterations, the result rises to 94% in simulation and 90% in the real world, with fixes including changes to pan grasp skills, grasp prompts, and placement offset (Chen et al., 6 Jul 2026).

The framework also functions as a front-end for other policy classes. Combining GaP’s modular skills, such as centering the wrist camera, before handing over to existing VLAs yields a 2x improvement in VLA success under severe position variation. Conversely, replacing the structured, multi-agent, graph-based system with single-agent monolithic code generation collapses success rates to zero because of type, reference, and interface errors and lack of modularity (Chen et al., 6 Jul 2026).

6. Terminological ambiguity and neighboring literatures

The acronym “VA” is heavily overloaded across arXiv literatures. In the robotics usage described above, it means Variational Automation. In other contexts it denotes different technical objects entirely.

Usage of “VA”	Meaning	Representative paper
Robotics manipulation	Variational Automation	(Chen et al., 6 Jul 2026)
Automotive human factors	vehicle automation	(Zhang et al., 29 Sep 2025)
Autonomous driving planning	Vision-to-Action	(Huang et al., 12 May 2026)
Inference, kinetics, and plasma physics	variational approximation, variational approach, or variational averaging	(Wingate et al., 2013, Korhonen et al., 2021, Wu et al., 2016, Possanner, 2017, Bieniasz et al., 2016)

This ambiguity matters because neighboring “VA” literatures address related but distinct automation problems. In automotive human-factors work, VA denotes vehicle automation systems, specifically Adaptive Cruise Control (ACC) and Lane Keeping Assist (LKA). A study of 36 participants with no prior VA experience compared owners’ manual, knowledge-based, and simulator-based training. All interventions improved post-training knowledge, but knowledge-based training had the strongest impact, produced significantly higher multiple-choice gains than owners’ manual training, and increased LKA and ACC activation frequency by 1.4 and 1.45 times, respectively. Older drivers maintained significantly longer LKA and ACC engagements once activated, with reported coefficients of 36.01 sec and 45.18 sec. The study concluded that short, targeted training can significantly improve safe and effective VA system use, particularly for senior drivers (Zhang et al., 29 Sep 2025). This suggests that, in road-vehicle contexts, safe automation under variation depends not only on controller capability but also on calibrated trust, boundary awareness, and standardized communication of corner cases.

In autonomous driving systems work, VA denotes Vision-to-Action models. These directly map visual sensor inputs to continuous action trajectories and are described as dominating in planning quality and real-time performance because of their streamlined design. A later Vision-Language-Action system, MindVLA-U1, argues that VLA has often trailed VA because semantic reasoning, temporal context, and continuous control have been combined through an inadequate interface. MindVLA-U1 reports 8.20 RFS versus 8.13 GT RFS on the long-tail WOD-E2E benchmark with 2 diffusion steps, and 16 FPS versus RAP-DINO’s 18 FPS while preserving natural-language interfaces (Huang et al., 12 May 2026). Here again, “VA” refers to a different automation axis than Variational Automation in robot manipulation.

Other disciplines use VA for non-automation concepts: automated variational inference in probabilistic programming (Wingate et al., 2013), extended variational approximations for GLLVMs (Korhonen et al., 2021), the variational approach of molecular kinetics (Wu et al., 2016), variational averaging in gyrokinetics (Possanner, 2017), and a Green’s-function variational approximation for orbital polarons in KCuF$\mathcal{G} = (V, E)$3 (Bieniasz et al., 2016). In encyclopedia usage, distinguishing these senses is essential, because only the robotics sense designates a task class centered on reliable automation under object and pose variation.