Grana Composes Subtask Progress to Track Long-Horizon Robot Manipulation

Grana Composes Subtask Progress to Track Long-Horizon Robot Manipulation

Zi-ang Cao, Kaiyuan Zheng, Yinzhen Xu, Haonan Duan, Jiasheng Gu +12 more

9 min read•Nov 19, 2026

Grana learns to select the unfinished subtask and predict its local progress, then combines these outputs into whole-task progress. The same language-conditioned model is pretrained on mostly short tasks and reused for longer compositions. Its progress targets account for both execution quality and the evaluated subtask’s position in the task sequence.

Predicting Progress Through Subtasks

Given camera observations, a task description, and the task’s subtask sequence, Grana selects the unfinished subtask and predicts its local progress. The supplied sequence specifies the user’s preferred execution order and evaluation granularity. Subtask descriptions are language inputs to a shared predictor; their order and number are not hard-coded into the architecture. The same task can therefore be queried with different valid orders or coarser or finer decompositions. Training or adaptation uses supervision defined under the chosen specification, while inference conditions on the supplied list. Whole-task progress measures completion relative to this user-specified procedure.

Six completed subtasks and halfway progress through the seventh give whole-task progress equal to the completed-subtask count plus local progress, divided by the total number of subtasks. If local progress falls to zero, the estimate becomes the completed-subtask count divided by the total number of subtasks. This composition assumes that preceding subtasks remain completed after a local failure; an incorrect selection can still distort the estimate.

In our model, subtask selection and progress prediction share a VLM that processes multiview observations sampled over time, providing context to distinguish a failed attempt from its recovery. A temporal transformer combines features for local-progress prediction through a shared head. Temporal aggregation can use later observations. In our policy-learning experiments, we score recorded trajectories offline.

Diagram of compositional progress prediction across subtasks

Supervision for Imperfect Execution

Selecting the right subtask does not reveal how much of it is complete. Elapsed time advances through both a successful insertion and repeated misalignment, yet these executions should receive different progress targets. We therefore combine subtask annotations with judgments of local action quality (AQ). Subtask annotations identify completed work. AQ judgments describe setbacks and recovery within the unfinished subtask.

Annotators mark subtask boundaries and intervals of suboptimal or regressed execution; unmarked intervals are nominal. Nominal execution follows the task normally. Suboptimal execution is degraded without physical failure, such as a misaligned screw that remains held. Regressed execution loses local progress through physical failure, such as dropping the screw. Separately, each subtask receives a binary outcome: 1 for successful completion and 0 for failure. These outcomes describe how the subtask ends: a completed subtask can still contain regressed intervals followed by recovery.

The annotations define a local progress target based on action quality and the subtask outcome without requiring numerical scores for individual frames. Within each uninterrupted attempt, we distribute progress according to action quality, giving nominal frames larger increments than suboptimal frames. In a successful subtask with retries, regression resets local progress to zero, and a subsequent attempt rebuilds it. The binary outcome anchors the final local-progress target.

Training Subtask Selection and Progress Prediction

Because Grana selects its own subtask, progress training must account for a description that is chosen incorrectly. For example, if insertion is only partly complete but the model selects fastening, fastening should receive zero progress rather than insertion’s partial progress. We use task order to define the target for each queried subtask: earlier completed subtasks receive one, the annotated unfinished subtask receives its local progress, and later subtasks receive zero.

We implement the shared predictor with Qwen3.5-4B and LoRA, training the adapters, temporal transformer, and progress head while keeping the base VLM frozen. Selection and progress use cross-entropy objectives. For progress training, scalar targets are encoded as distributions over neighboring bins; the head predicts a distribution whose mean gives a continuous progress estimate. Both objectives place greater weight on predictions farther from the target. Pretraining and adaptation retain the same architecture and language interface. At inference, the model receives observations, the description, and the subtask list, without annotation inputs.

Using Progress for Policy Learning

Progress differences provide feedback about recorded actions: an increase indicates advancement, while a decrease indicates a setback. We use these differences in two ways, holding the conversion procedure fixed across reward models in each comparison.

Experiments

Our experiments address four questions:

  • Supervision: Can structured subtask and quality annotations provide practical and useful progress supervision?
  • Generalization: How well does Grana generalize across tasks, horizons, embodiments, and datasets?
  • Pretraining and composition: How does reward-model pretraining improve long-horizon adaptation, and does subtask composition improve progress prediction in both Grana and a frontier VLM such as GPT-6 Astra?
  • Policy learning: Do the resulting progress estimates improve downstream policy learning?

Experimental Setup

Policy experiments use the GR00T-2 world-action model, holding initialization, data, and learning procedure fixed within each task and comparison while varying reward models.

Reward Prediction and Generalization

We evaluate zero-shot prediction on held-out tasks, adaptation to longer tasks and a different embodiment, and whole-trajectory ranking on a public benchmark.

On the 50 held-out tasks without target-task reward labels or fine-tuning, Grana achieves RMSE 0.180 versus 0.227 for the strongest baseline, RoboMeter, a 20.7% reduction. Increasing pretraining from 15 to 200 tasks reduces held-out RMSE by 35.3%.

Five long-horizon robot manipulation tasks used in the experiments

Subtask Composition and Progress Prediction

We test whether subtask composition improves progress prediction, then examine the roles of subtask selection and local-progress prediction.

On airplane model assembly, direct whole-task and composed predictors start from base Qwen without reward-model pretraining and use matched adaptation budgets. Automatic subtask selection and local-progress composition reduce RMSE by 17.3% with D15 and 13.4% with D100, including selection errors.

In zero-shot car model assembly, composition reduces accumulated RMSE from 1.578 to 0.524 for GPT-6 and by 32.4% for base Qwen3.5-4B. These gains show that our proposed subtask composition improves progress prediction in both models.

Downstream Policy Improvement

We test whether predicted progress improves policies through offline reinforcement learning and progress-weighted behavior cloning. Appendix E additionally reports test-time model predictive control on tractor assembly, where Grana decides when to plan and selects among imagined futures.

Each reward source scores the same BC-policy rollouts on two held-out tasks: 50 for star insertion (one subtask) and 45 for tube transfer from a box to a rack (three subtasks). We convert predicted progress into rewards using the same rule across all reward models. These rewards supervise IQL critic learning, followed by BPPO actor updates. Zero-shot means reward models receive no target-task reward annotations; policies train on scored target-task rollouts.

Screw insertion example illustrating local progress during manipulation

Acknowledgments

We used AI, including GPT-6 Astra, Claude Fable 5, and Opus 5, to polish the language, improve the presentation clarity of all figures, and assist with code implementation.

Parameter Selection and a Worked Example

The airplane assembly insertion example follows subtask 7 of 9. Six subtasks are already complete, so the whole-task target is the completed-subtask count plus local progress, divided by the total number of subtasks. The example contains 1,117 synchronized samples at 30 frames per second, indexed from 0 through 1,116. The first attempt is nominal through frame 249; frames 250–820 are regressed; frames 821–1,116 show nominal recovery to a successful outcome.

Progress-Consistency Metrics

The airplane and car model test episodes are scored with whole-task progress, calculated as accumulated subtask progress divided by the total number of subtasks and averaged over the four camera views, at about 1 Hz. Both metrics use the human annotations of these episodes.

Annotators mark keyframes of high- and low-quality execution within each subtask. For every pair of a high-quality and a low-quality keyframe of the same subtask from different episodes, we take the score at the evaluation point nearest to each keyframe (within 0.5 s). Q-Rank is the percentage of pairs in which the high-quality keyframe scores strictly higher. Independent random scores drawn from the same continuous distribution rank either frame higher with equal probability, giving Q-Rank 50 on average. The pairs involve 12 low-quality keyframes per task.

Training and Inference

Pretraining and adaptation use the objectives described above, retaining the same architecture and language interface. During adaptation, descriptions selected by the model itself receive the order-conditioned targets.

Inference uses two sequential calls: the first selects a subtask from the supplied ordered list, and the second predicts local progress conditioned on that subtask’s description. The selected index and local prediction are composed using the whole-task progress rule. Inference requires observations and text inputs, without subtask-boundary, AQ, or outcome annotations.

Switching MPC On and Off

  • Progress stall. Whole-task progress has not grown over the last specified time interval: the policy is stuck, or repeated attempts are failing.
  • Subtask jump. The smoothed subtask index skips ahead: the policy has moved to a later step before completing the current one, for example reaching for the drill before mounting a wheel.

Tractor Assembly

Without MPC, the policy completes 2 of the 10 assemblies; with MPC, it completes 9 of 10. Failures without MPC include dropped or tilted wheels and skipped steps, such as reaching for the drill before mounting a wheel. In the remaining MPC failure, all four wheels are mounted, but the tractor is not set upright. These results indicate that Grana’s subtask and progress estimates can both decide when planning is needed and guide action selection, without any policy update.

Frequently Asked Questions

How does Grana estimate whole-task progress? It selects the unfinished subtask, predicts its local progress, and combines that estimate with the task’s ordered subtask list.

How does Grana distinguish successful progress from setbacks? Subtask annotations identify completed work, while action-quality annotations mark nominal, suboptimal, and regressed execution. Regression can reset local progress before a subsequent attempt rebuilds it.

What does Grana receive at inference time? It receives observations, a task description, and the supplied subtask list. It does not require subtask-boundary, action-quality, or outcome annotations at inference.

How did MPC affect tractor assembly results? The policy completed 2 of 10 assemblies without MPC and 9 of 10 with MPC, without any policy update.