Capek 0.5 | XPENG Robotics

An execution-centric vision-language model for embodied intelligence.

Robot execution is iterative: each action changes what must be perceived, reasoned about, grounded, and verified next. Capek 0.5 organizes post-training around four execution-facing capabilities and consolidates their specialists into a single inference-time model.

arXiv

35B-A3B domain means

Capek 0.5 leads across execution-facing capabilities.

70.7 63.6 57.0 57.9 66.2 66.3

Spatial

62.7 53.7 41.7 53.4 52.2 48.0

Temporal

68.1 51.0 53.5 62.7 55.2 59.9

Guidance

61.5 53.4 47.9 46.3 49.4 46.6

State

82.9 62.9 66.4 63.3 71.8 72.8

General

Domain means are replotted from the paper figure with the lower axis cropped at 30 to emphasize score separation in the release poster.

Capek-StateBench

A held-out benchmark for verifying execution state and task progress.

PSV / Video

Verify support relation.

Question: Can we verify that the trash can is on the floor? GT: yes.

PSV / Video

Judge final object state.

Question: Is the bottom cabinet door open or closed? GT: closed.

StateBench-T task-condition progress example

TSV / Task condition

Estimate task progress.

Question: Sort vegetables into three mixing bowls by type. GT: {"progress": 46.15, "next_action": "pick up the leek from the wicker basket"}.

StateBench-T primitive-skill progress example

TSV / Primitive skill

Predict the next action.

Question: Pour diced pineapple into the bowl, then return the cutting board to the sink. GT: {"progress": 50, "next_action": "pour the diced pineapple from the cutting board into the bowl"}.

StateBench data construction from execution checkpoints — State Verification examples are anchored at primitive, skill, and task checkpoints. The model sees only causal visual evidence at the selected checkpoint.

Why this benchmark

Most public suites do not ask whether execution actually reached the required state.

Capek-StateBench targets the verification side of embodied execution: observable physical predicates, task progress, and the immediate next action needed to continue. It is held out from specialist training and aligned with the State Verification objective.

StateBench-P

Physical-state verification

500 examples evaluate object states and relations from a single image, short clip, or key-frame set. Predictions are normalized and scored by exact match.

StateBench-T

Task-state verification

500 examples evaluate progress and next-step prediction from a causal visual prefix: 213 task-condition records and 287 primitive-skill records.

Scoring

Progress + action

Physical scores use exact match. Task scores combine progress accuracy and next-action semantic matching:

s_p = max ( 0, 1 - |p̂ - p| 25 )

s_T = 5s_p + 4s_a 9

s_p: progress score; p̂: predicted progress; p: reference progress; s_a: next-action semantic exact match; s_T: task-track score.

Static QA samples

Single-turn capability checks with structured outputs.

State Verification

Physical-state verification

Question

Based on observations, is the lighter on or off?

Answer

off

The model tracks the lighter across chronological observations and returns the final state label.

Execution-centric taxonomy and method

Capek 0.5 turns embodied execution into trainable, verifiable capabilities.

Execution-centric capability overview — **Execution-centric capability view.** Spatial, Temporal, Guidance, and State requirements recur across an observation-reasoning-action loop rather than forming mutually exclusive stages.

Why this taxonomy

Execution is a loop, so training has to cover the loop.

A robot does not just answer what is in an image. It must localize objects, understand what changed, choose grounded actions, and verify whether the world reached the required state after acting.

Capek 0.5 therefore organizes post-training around four capabilities that repeatedly appear during execution, then turns each capability into structured, checkable supervision.

Representative data contracts — Representative contracts for verifiable supervision.

See where execution can happen

Spatial reasoning handles geometry, object relations, viewpoint changes, distance, depth, and placement feasibility.

Track what changes over time

Temporal understanding handles event ordering, temporal grounding, long videos, and observation-to-observation changes.

Turn intent into action targets

Action guidance grounds language into referents, points, affordance regions, and ordered image-space trajectories.

Verify whether execution succeeded

State verification checks physical predicates, progress, next actions, and task-state completion.

Make outputs verifiable

Records are organized by execution role and output contract.

Each capability uses output formats that can be parsed and scored: choices for spatial judgments, time intervals for events, coordinates and trajectories for guidance, and JSON-style predicates or progress fields for state verification.

This lets each specialist learn from feedback that matches the thing execution actually needs: a location, a temporal relation, an action target, or a state decision.

State candidates: 20,000 PSV rows; 24,195 PVE progress-value-estimation rows. Counts describe the candidate inventory, not the run-specific training mixture or exposure.

Spatial · Perception & reasoning 84.0K · 26.5%

Direction / orientation 20,730 · 24.7%

Multi- / cross-view 20,194 · 24.0%

Relations / metric 20,499 · 24.4%

Navigation / scene 14,028 · 16.7%

Other verified tasks 8,576 · 10.2%

Temporal · Events & change 92.6K · 29.2%

Structured video reasoning 55,531 · 60.0%

Temporal / ST grounding 12,594 · 13.6%

Open video reasoning 24,441 · 26.4%

Guidance · Targets & paths 95.9K · 30.3%

Grounding 38,767 · 40.4%

Affordance 24,408 · 25.4%

Trajectory 11,277 · 11.8%

Task understanding 21,496 · 22.4%

State · Physical & task state 44.2K · 14.0%

PSV · Physical-state verification 20,000 · 45.3%

PVE · Progress-value estimation 24,195 · 54.7%

Capability-oriented data inventory grouped by execution role.

Specialize during training

Four same-origin specialists acquire focused behaviors.

Spatial, Temporal, Guidance, and State specialists start from the same backbone but receive capability-specific data, parsers, and reward geometry. The separation keeps feedback precise while capabilities are being acquired.

Spatial specialist

Geometry and relations

Trains on relations, metric reasoning, multi-view reasoning, and scene-level spatial judgments.

Temporal specialist

Events over time

Learns event ordering, temporal localization, video QA, and long-video reasoning.

Guidance specialist

Targets and trajectories

Optimizes boxes, points, regions, affordances, and ordered image-space paths.

State specialist

Predicates and progress

Checks state labels, binary predicates, progress values, next actions, and JSON validity.

Specialist consolidation pipeline for Capek 0.5

Merge for deployment

Specialists are folded back into one inference-time model.

TIES initializes a unified student in weight space, then routed MOPD consolidates student-generated prefixes with the responsible frozen specialist. The deployed checkpoint exposes one autoregressive interface while retaining behaviors acquired by focused specialists.

Takeaway 1

Separate capability acquisition

Train four same-origin specialists independently with capability-aligned data, objectives, and rewards.

Takeaway 2

Initialize with TIES

Merge specialist task vectors in weight space to initialize a unified student checkpoint.

Takeaway 3

Consolidate with routed MOPD

Route student-generated prefixes to the responsible frozen specialist during training; inference uses one autoregressive model.

Core benchmark leaderboard

Execution-facing benchmark results across capability families and model scales.

The default view keeps 15 high-signal rows; switch to all benchmarks for the full matrix. Metrics are shown next to benchmark names; the Capek column is pinned first, with best scores bolded and second-best scores underlined.

35B-A3B agentic evaluation

Embodied agent evaluation in simulated environments.

Model	EB-HAB Avg.	EB-ALF Avg.
Embodied-R1.5 8B	21.7	14.7
RynnBrain1.1-9B	44.7	35.3
RoboBrain2.0-32B	38.3	25.0
RynnBrain-30B-A3B	26.7	26.3
Qwen3.6-35B-A3B	46.0	50.7
Capek0.5 35B-A3B	63.0	55.3

Model	Score	W	B
Embodied-R1.5 8B	31.7	33.4	31.7
RynnBrain1.1-9B	29.1	34.3	23.4
RoboBrain2.0-32B	38.1	36.2	34.3
RynnBrain-30B-A3B	17.2	22.2	13.5
Qwen3.6-35B-A3B	30.2	37.4	28.8
Capek0.5 35B-A3B	36.4	38.6	32.2

Qualitative samples

Sample rollouts across EmbodiedBench and VIGIL.

EmbodiedBench sample

Long-horizon manipulation with 20 successful steps.

EmbodiedBench / ALF episode 10

Start

Question

Pick up knife, slice apple, put knife in bowl, heat slice of apple in microwave, put apple slice on table.

instruction issued

The rollout starts from the task instruction before any environment-changing action.

EmbodiedBench sample

Goal-conditioned object retrieval in Habitat.

Selected EmbodiedBench Habitat rollout frame

EmbodiedBench / Habitat visual episode 8

Start

Question

Retrieve a purple fruit and place it in the sink.

instruction issued

The rollout starts with a target object and destination but no executed action.

VIGIL sample

Closed-loop execution with step-level action selection.

VIGIL / sample-627

Step 1

Question

Find the toaster, turn it on, then report success.

navigate forward, 2

The agent moves toward the table; no target interaction has been credited yet.

VIGIL sample

Sequential rearrangement with terminal success reporting.

VIGIL / sample-813

Start

Question

Pick up the laptop, place it on the coffee table, then report success.

instruction issued

The laptop is still on the armchair before any environment-changing action.

Citation

@misc{chen2026capek05executioncentricvisionlanguage,
  title={Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence},
  author={Ying Chen and Weizhen Li and Zhe Hu and Zhenjiang Li and Rui Jiang and Zhifeng Gu and Lihuang Fang and Jiangping Liu and Lei Yi and Jie Chen},
  year={2026},
  eprint={2608.06756},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2608.06756},
}

Capek 0.5 leads across execution-facing capabilities.

A held-out benchmark for verifying execution state and task progress.

Verify support relation.

Judge final object state.

Estimate task progress.

Predict the next action.

Most public suites do not ask whether execution actually reached the required state.

Physical-state verification

Task-state verification

Progress + action

Single-turn capability checks with structured outputs.

Physical-state verification

Capek 0.5 turns embodied execution into trainable, verifiable capabilities.

Execution is a loop, so training has to cover the loop.

See where execution can happen

Track what changes over time

Turn intent into action targets

Verify whether execution succeeded

Records are organized by execution role and output contract.

Spatial · Perception & reasoning 84.0K · 26.5%

Temporal · Events & change 92.6K · 29.2%

Guidance · Targets & paths 95.9K · 30.3%

State · Physical & task state 44.2K · 14.0%

Four same-origin specialists acquire focused behaviors.

Geometry and relations

Events over time

Targets and trajectories

Predicates and progress

Specialists are folded back into one inference-time model.

Separate capability acquisition

Initialize with TIES

Consolidate with routed MOPD

Execution-facing benchmark results across capability families and model scales.

35B-A3B core benchmark leaderboard

Embodied agent evaluation in simulated environments.

Average task success on EB-HAB and EB-ALF

Primary score, world completion, and benchmark success

Sample rollouts across EmbodiedBench and VIGIL.

Long-horizon manipulation with 20 successful steps.

Start

Goal-conditioned object retrieval in Habitat.

Start

Closed-loop execution with step-level action selection.

Step 1

Sequential rearrangement with terminal success reporting.

Start

Citation