SeqVLA: Completion-Aware Robotic Manipulation

• SeqVLA is a completion-aware extension of π0 that detects subtask completion and autonomously triggers transitions for reliable multi-stage manipulation.
• The architecture employs a dual-head design combining a SigLIP vision encoder, Gemma language and action experts, and a lightweight binary classifier for effective control.
• Empirical results on salad and candy packing tasks demonstrate that joint finetuning significantly reduces sequencing errors and enhances long-horizon success rates.

SeqVLA is a completion-aware extension of $\pi_0$ for long-horizon robotic manipulation, designed for tasks composed of multiple interdependent subtasks that must be executed in strict sequence (Yang et al., 17 Sep 2025). The model augments the base Vision-Language-Action architecture with a lightweight detection head that perceives whether the current subtask is complete, enabling it both to generate continuous low-level manipulation actions and to autonomously trigger transitions between subtasks. In the reported instantiation, SeqVLA is built on $\pi_0$ with a SigLIP vision encoder, a Gemma-2B language backbone, and a Gemma-300M action expert, and is evaluated on salad packing and candy packing as multi-stage bimanual manipulation tasks (Yang et al., 17 Sep 2025).

1. Problem formulation and motivation

Long-horizon manipulation commonly requires a robot to execute multiple subtasks in a fixed order, where switching too early, lingering too long, or repeating a completed step can induce downstream failures. SeqVLA is motivated by the observation that standard VLA models such as $\pi_0$ are strong at continuous, low-level control but lack an internal, learned signal for subtask completion. When naively chained for sequential tasks, $\pi_0$ tends to switch prematurely or stall due to no progress-monitoring signal, repeat completed subtasks or execute them out of order because visually similar scenes occur across stages, and propagate early-stage errors into later subtasks (Yang et al., 17 Sep 2025).

The central premise of SeqVLA is that subtask completion detection should gate when the controller advances to the next prompt or goal. Each subtask is paired with a unique language prompt, and the model operates under a known, strict sequence of these prompts. Within that setting, SeqVLA couples action generation with a learned completion detector so that stage transitions are decided from the same multimodal context used to generate actions. This design directly targets sequence brittleness rather than low-level control alone.

2. Architectural design

SeqVLA is built on $\pi_0$, reusing its pretrained multimodal backbone and action expert for continuous control (Yang et al., 17 Sep 2025). Its inputs are three RGB camera streams, robot joint states, and a language prompt describing the current subtask goal. The three cameras are a front/top global view, a left wrist camera, and a right wrist camera. The robot platform is a dual-arm Aloha system, and the action output is continuous 14-DoF low-level control for bimanual manipulation.

The defining architectural modification is a dual-head output structure. The action head performs the standard control role inherited from $\pi_0$. The completion detection head is a lightweight binary classifier that infers whether the current subtask is complete. It shares representations with the action pathway by consuming the 1024-d feature $F$ from the action expert, an arrangement intended to preserve computational efficiency while aligning detection with action-relevant state.

The completion head is specified as

$p = \sigma(W \cdot F + b),$

with $W \in \mathbb{R}^{1024}$, $b \in \mathbb{R}$, and

$\pi_0$0

Here $\pi_0$1 is the task state indicator. In the execution logic used by the paper, $\pi_0$2 close to 1 indicates continuation of the current subtask, whereas $\pi_0$3 close to 0 indicates termination and transition to the next subtask. The architecture therefore embeds completion awareness directly into the policy’s forward pass rather than delegating it to an external controller (Yang et al., 17 Sep 2025).

3. Training objectives and finetuning regimes

The action component follows $\pi_0$4’s flow-matching formulation. For a noisy action $\pi_0$5 and ground-truth action $\pi_0$6, SeqVLA samples $\pi_0$7 and interpolates

$\pi_0$8

The policy then predicts the target velocity $\pi_0$9 through the loss

$\pi_0$0

The paper refers to $\pi_0$1 as the standard $\pi_0$2 action loss provided by the 300M action expert. Completion detection is trained with binary cross-entropy,

$\pi_0$3

and the combined objective is

$\pi_0$4

with $\pi_0$5 (Yang et al., 17 Sep 2025).

Two training axes are investigated: optimization schedule and backbone freezing. Joint finetuning trains action and detection heads simultaneously, whereas sequential finetuning first trains the action head and backbone while the detection head is frozen, then freezes the action head and backbone to finetune the detection head. The backbone can either remain unfrozen, allowing full finetuning of the vision-language stack, or be frozen so that only the action expert and heads are trained.

Configuration	Optimization schedule	Backbone
SeqVLA-J	Joint finetuning	Unfrozen
SeqVLA-JF	Joint finetuning	Frozen
SeqVLA-S	Sequential finetuning	Unfrozen
SeqVLA-SF	Sequential finetuning	Frozen

The paper states that, in joint finetuning, gradient flow couples the two heads via shared features from the action expert, improving representation alignment. A plausible implication is that completion cues are learned in a form more directly useful for transition decisions when they are optimized together with manipulation control.

4. Sequential controller and execution semantics

At inference time, SeqVLA emits both an action and a completion probability $\pi_0$6 (Yang et al., 17 Sep 2025). The controller uses a threshold $\pi_0$7 to determine when to advance. Empirically, $\pi_0$8 works best. If $\pi_0$9, the controller immediately stops the current action, sends the robot to a home pose for consistent re-initialization, switches to the next subtask’s language prompt, and continues execution.

The paper characterizes threshold selection as a trade-off. A lower threshold such as $\pi_0$0 is overly conservative because stop signals become rare. A higher threshold is overly sensitive and can cause premature stopping and oscillation. The reported controller is therefore a simple thresholded state-transition mechanism layered on top of a continuous policy, but its transition signal is learned rather than manually engineered.

A theoretical perspective on cascade failures is also given. If subtask $\pi_0$1 has completion detection error rate $\pi_0$2, then the probability of at least one stage-wise failure over $\pi_0$3 subtasks is

$\pi_0$4

This formulation makes explicit why better completion detection can disproportionately improve long-horizon execution: reducing each $\pi_0$5 lowers the overall failure probability multiplicatively. The paper notes an optional stabilization strategy requiring multiple consecutive timesteps with $\pi_0$6 before advancing, but this is not used in the reported experiments (Yang et al., 17 Sep 2025).

5. Experimental setting, data, and tasks

The empirical study uses a bimanual 14-DoF Aloha robot instrumented with three cameras: a front/top global camera and left and right wrist cameras (Yang et al., 17 Sep 2025). Training data are collected by teleoperation and include both subtask-level demonstrations and full long-horizon demonstrations. The subtask-level demonstrations are used to finetune SeqVLA’s action and detection components; full long-horizon demonstrations are used to finetune the baseline $\pi_0$7 for end-to-end execution.

Completion labels are binary at the frame level and indicate whether an action contributes to the current subtask or is post-completion/reset behavior. The typical pattern is described as $\pi_0$8. The completion criteria are strict: the object must be placed successfully, the gripper must be fully released, and there must be no gripper-container contact. This labeling policy is intended to ensure that reset motions do not count toward the subtask.

Two multi-stage tasks are reported. Salad packing contains seven subtasks in strict order: load spinach, load coleslaw, load meatball, load chicken, load tomato, add sauce cup, and close container. Candy packing contains four subtasks including repetitions: pick gummies, place Kinder chocolate twice, add Snickers twice, and place lollipop. The dataset includes 50 demonstrations per subtask, yielding 350 subtask episodes for salad and 200 subtask episodes for candy (Yang et al., 17 Sep 2025).

Evaluation uses overall long-horizon success rate as the primary task metric. Completion detection is analyzed using entropy of classification outputs, where lower values indicate more confident predictions, and the Kolmogorov–Smirnov statistic between execution-phase and completion-phase output distributions, where higher values indicate better separation. All reported KS p-values are below 0.001.

6. Empirical findings

The main empirical pattern is consistent across tasks and ablations: allowing the pretrained backbone to adapt yields higher subtask success rates than freezing it, and among the unfrozen variants, joint finetuning produces more decisive and reliable completion predictions than sequential finetuning (Yang et al., 17 Sep 2025). This is reflected in both entropy and KS statistics.

For salad packing, overall entropy improves from 1.35 for SeqVLA-S to 0.76 for SeqVLA-J, with overall KS approximately 0.75. Per-subtask examples include spinach from 1.32 to 0.76 with KS 0.77, coleslaw from 1.28 to 0.70 with KS 0.79, sauce from 1.51 to 0.57 with KS 0.81, and container closing from 1.19 to 0.48 with KS 0.79. For candy packing, overall entropy improves from 1.11 for SeqVLA-S to 0.79 for SeqVLA-J, with overall KS approximately 0.72. Per-subtask examples include Snickers from 1.78 to 0.67 with KS 0.85, Kinder from 1.10 to 0.79 with KS 0.72, and lollipop from 0.83 to 0.67 with KS 0.80.

The comparison against the baseline $\pi_0$9 is particularly informative. The paper states that $\pi_0$0 finetuned end-to-end on full sequences achieves comparable raw manipulation success on individual subtasks but frequently fails at sequencing: it executes out of order, repeats subtasks, and cannot maintain the designed long-horizon structure. By contrast, SeqVLA-J eliminates sequence-related failures by using the completion detection head to trigger transitions. When failures occur, they stem from genuine manipulation difficulty rather than incorrect ordering. The summary plot shows SeqVLA-J achieving higher overall long-horizon success rates than $\pi_0$1, although exact percentages are not reported. This suggests that completion awareness is a principal determinant of long-horizon robustness in the reported setting.

7. Positioning, limitations, and nomenclature

SeqVLA is positioned as a modification of $\pi_0$2 rather than a departure from the VLA paradigm (Yang et al., 17 Sep 2025). Relative to prior progress monitors and failure detectors, the paper emphasizes that SeqVLA does not keep termination signals separate from action generation; instead, the integrated dual-head design uses a shared representation for action and detection. Relative to hierarchical RL, options, planner+policy systems, and mixture-of-experts approaches, SeqVLA remains within a single VLA model and learns when to advance subtasks from the same multimodal context used to generate actions. The paper also distinguishes it from other VLA families such as RT-1/2, OpenVLA, VIMA, CoT-VLA, and SmoVLA, noting that many rely on tokenized actions or latent trajectories and external sequencing logic, whereas SeqVLA builds completion-aware action generation directly on $\pi_0$3’s continuous flow model.

The reported limitations are equally specific. SeqVLA assumes a strict, known subtask sequence, one binary detector for the current subtask state, and a single active task prompt at a time. It is evaluated on two real, structured domains; generalization to unseen sequences, branching, or concurrent subtasks is left for future work. The method requires subtask-level demonstrations with framewise completion labels, and although joint finetuning improves confidence substantially, misdetections can still produce early or late transitions. The paper notes that debouncing or temporal smoothing could help, but these are not used in the reported experiments. Latency under different hardware and control rates is not reported.

A nomenclature issue is worth separating from the manipulation method itself. A distinct paper, "SeqVLM: Proposal-Guided Multi-View Sequences Reasoning via VLM for Zero-Shot 3D Visual Grounding," addresses zero-shot 3D visual grounding rather than long-horizon manipulation, and the query term “SeqVLA” is explicitly noted there as appearing to refer to SeqVLM in some contexts (Lin et al., 28 Aug 2025). In current usage, however, SeqVLA properly denotes the completion-aware sequential execution model for long-horizon manipulation introduced as an extension of $\pi_0$4 (Yang et al., 17 Sep 2025).