cs.CLOct 26, 2026
SaveOmniBrain-0: A Unified Agentic Model for Embodied Intelligence
Organizations: University of Macau × Tsinghua University · University of Macau, Tsinghua University
Abstract
General-purpose embodied intelligence requires robots to coordinate diverse skills over long horizons, track task progress, and recover from failures. Specialist policies alone do not ensure task-level reliability, while powerful multimodal agents rely on costly reasoning and external experience that compact models struggle to reproduce. We introduce OmniBrain-0, a lightweight unified embodied agent that coordinates manipulation and navigation through a shared high-level decision model, integrating task understanding, planning, progress tracking, and feedback-driven recovery. Our progressive training framework combines embodied pretraining, domain-specialized agent distillation, and multi-teacher on-policy distillation (MOPD). It distills powerful teachers' interaction experience into manipulation and navigation experts, then consolidates their decision capabilities into OmniBrain-0. Unification therefore occurs within the learned decision model rather than only at the tool interface. We further introduce Rubric-Guided Recursive Self-Improvement (RSI), a data flywheel in which a Meta-Agent turns execution evidence into refined rubrics, step-level reminders, and verified trajectories for domain-expert updates before their reconsolidation into OmniBrain-0 through MOPD. We evaluate OmniBrain-0 on LIBERO-Pro, LIBERO-Plus, RoboCasa365, RoboTwin 2.0, and R2R-CE Val-Unseen, as well as multiple real-world manipulation tasks. OmniBrain-0 combines strong cross-domain task performance with substantially lower inference and execution overhead than its closed-source teacher, demonstrating efficient transfer and consolidation of embodied capabilities into a compact unified agent.
Figures & tables
Figure 1: Cross-domain embodied-agent performance. Success rates on LIBERO-Pro, RoboCasa365, RoboTwin 2.0, and R2R-CE Val-Unseen, comparing specialist policies, few-shot teachers, and OmniBrain-0.
Figure 2: Overview of OmniBrain-0. Top: Three-stage training: embodied pretraining, domain-specialized distillation into manipulation and navigation experts, and multi-expert consolidation (MOPD) into a single compact agent. Middle: Given a task instruction, observations, and an optional experience library, OmniBrain-0 understands, plans, orchestrates skills, and recovers from failures across manipulation and navigation. Bottom: Rubric-Guided Recursive Self-Improvement (RSI) uses interaction evidence to refine step-level reminders and curate verified trajectories, then updates domain experts and consolidates their capabilities through MOPD for the next interaction round. The radar chart reports success rates (%) on five benchmarks, comparing OmniBrain-0 with the baseline policy ( for LIBERO-Pro, LIBERO-Plus, RoboTwin 2.0, and RoboCasa365; InternVLA-N1 for R2R-CE Val-Unseen) and the few-shot teacher GPT-5.6-Sol.
Figure 3: Agentic execution framework and task execution examples. Left: Unified agentic system design comprising general perception tools, VLA manipulation tools, VLN navigation tools, and an optional experience library. Center: OmniBrain-0 performs observation-driven reasoning and tool orchestration through closed-loop interaction with task environments, using execution feedback to guide subsequent decisions. Right: Representative manipulation and navigation examples illustrating step-by-step reasoning, tool invocation, and task progression.
Figure 4: Interaction data construction. Raw interaction logs, images, and environment states pass through interaction normalization, trajectory quality control, reasoning reconstruction, and consistency validation. Canonical successes and successful recoveries form the training data; incomplete, unsuccessful, noisy, evidence-deficient, or inconsistent trajectories are discarded.
Figure 5: Three-stage training framework of OmniBrain-0. Left: Embodied VQA and agent interaction data establish the shared OmniBrain-Base initialization. Middle: Curated expert trajectories, including canonical successes and successful recoveries, supervise domain-specialized manipulation and navigation agents. Right: MOPD consolidates the frozen domain experts into a unified student through masked reverse-KL supervision on student-generated decision prefixes, updating only the student.
Figure 6: Rubric-based policy update and reminder-policy validation. Left: A verifier assesses collected successful and failed trajectories. The Meta-Agent analyzes execution evidence and adds, revises, or reinforces robotics rubrics supported by trajectory patches. Right: Rubric candidates inform the reminder policy, which conditions OmniBrain-0 together with the task input while model parameters remain fixed. Verified evaluation trajectories provide feedback for the next round, and effective reminders are retained as valid reminders.
Figure 7: RSI data flywheel. OmniBrain-0 collects challenging cases that inform the rubric-guided reminder policy and trajectory curation. The Meta-Agent coordinates analysis, guidance, curation, and improvement, with task relevance, diversity, learning value, and difficulty considered during selection. Verified trajectories form manipulation and navigation pools for expert training. MOPD consolidates the updated experts into the unified model, which is redeployed for the next collection round.
| Method | Success Rate (%) | Time | ||||||||
| Spat-T | Spat-S | Obj-T | Obj-S | Goal-T | Goal-S | L10-T | L10-S | Overall | ||
| Training-Based Skill Policies | ||||||||||
| OpenVLA [ 22 ] | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | / |
| NORA [ 17 ] | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | / |
| [ 5 ] | 0.0 | 0.0 | 0.0 | 2.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.3 | / |
| MolmoAct [ 25 ] | 0.0 | 0.0 | 0.0 | 6.0 | 0.0 | 0.0 | 6.0 | 0.0 | 1.5 | / |
Table 1: Success rates (%) on LIBERO-Pro under Task (T) and Swap (S) perturbations across four suites.
| Method | Success Rate (%) | Time | |||||||
| Camera | Robot | Language | Light | Back- ground | Noise | Layout | Overall | ||
| Training-Based Skill Policies | |||||||||
| OpenVLA [ 22 ] | 0.8 | 3.5 | 23.0 | 8.1 | 34.8 | 15.2 | 28.5 | 15.6 | / |
| NORA [ 17 ] | 2.2 | 37.0 | 65.1 | 45.7 | 58.6 | 12.8 | 62.1 | 39.0 | / |
| [ 5 ] | 13.8 | 6.0 | 58.8 | 85.0 | 81.4 | 79.0 | 68.9 | 53.6 | / |
| OpenVLA-OFT [ 21 ] | 56.4 | 31.9 | 79.5 | 88.7 | 93.3 | 75.8 | 74.2 | 69.6 | / |
Table 2: Success rates (%) on LIBERO-Plus across seven perturbation types.
| Method | Success Rate (%) | Time | |||
| Atomic-Seen | Composite-Seen | Composite-Unseen | Overall | ||
| Training-Based Skill Policies | |||||
| Diffusion Policy [ 10 ] | 15.7 | 0.2 | 1.3 | 6.1 | / |
| [ 5 ] | 34.6 | 6.1 | 1.1 | 14.8 | / |
| [ 4 ] | 39.6 | 7.1 | 1.2 | 16.9 | / |
| Giga World-Policy 0.1 [ 54 ] | 44.4 | 11.8 | 2.9 | 20.7 | / |
Table 3: Success rates (%) on RoboCasa365 across seen and unseen task splits.
| Method | Success Rate (%) | Time | ||
| Clean | Randomized | Overall | ||
| Training-Based Skill Policies | ||||
| X-VLA [ 64 ] | 72.8 | 41.2 | 57.0 | / |
| [ 4 ] | 81.2 | 75.6 | 78.4 | / |
| LingBot-VLA [ 49 ] | 86.0 | 83.6 | 84.8 | 2.8 |
| ABot-M0 [ 53 ] | 86.06 | 85.08 | 85.57 | / |
Table 4: Success rates (%) on RoboTwin 2.0 under Clean and Randomized settings.
| Method | SR | SPL | OSR | NE | Time |
| Training-Based Skill Policies | |||||
| NaVid [ 59 ] | 37.4 | 35.9 | 49.1 | 5.47 | / |
| Uni-NaVid [ 58 ] | 47.0 | 42.7 | 53.3 | 5.58 | / |
| AO-Planner [ 6 ] | 47.0 | 33.0 | 59.0 | 5.55 | / |
| ActiveVLN [ 62 ] | 52.9 | 45.7 | 62.1 | 5.34 | / |
| StreamVLN [ 48 ] | 56.9 | 51.9 | 64.2 | 4.98 | / |
Table 5: R2R-CE Val-Unseen results. Skill policies use the full split; agents use the 100-episode rand100 subset. SR, SPL, and OSR are percentages; NE is in meters.
| Method | Accuracy (%) | |||
| MP3D | AI2-THOR | ScanNet | Overall | |
| InternVL3-8B [ 69 ] | 76.3 | 74.2 | 78.5 | 76.3 |
| Cambrian-S-7B [ 52 ] | 71.3 | 72.9 | 74.3 | 72.8 |
| Qwen3.5-4B | 77.7 | 71.8 | 77.1 | 75.5 |
| Qwen3.5-9B | 78.6 | 72.1 | 81.2 | 77.2 |
| OmniBrain-0 (4B) | 81.1 | 73.0 | 83.2 | 79.0 |
Table 6: Spatial question-answering accuracy (%) on EmbSpatial-Bench across three source environments.
| Method | Accuracy (%) | |||
| Location | Placement | Unseen | Overall | |
| SpaceLLaVA [ 39 ] | 5.8 | 4.3 | 4.0 | 4.8 |
| RoboPoint [ 55 ] | 22.9 | 9.3 | 8.4 | 13.9 |
| Molmo-7B [ 11 ] | 21.9 | 12.9 | 12.2 | 15.9 |
| RoboRefer-2B-RFT [ 65 ] | 52.0 | 54.0 | 41.6 | 49.8 |
| Qwen3.5-4B | 46.0 | 20.0 | 26.0 | 31.1 |
Table 7: Spatial referring accuracy (%) on RefSpatial-Bench: object localization (Location), free-space selection (Placement), and novel relation combinations (Unseen).
Figure 8: Real-world manipulation performance. Comparison of the standalone baseline , the compact unified agent OmniBrain-0, and GPT-5.6-Sol (few-shot). Success rates are reported as percentages.
| Method | LIBERO-Pro | LIBERO-Plus | RoboCasa365 | RoboTwin 2.0 | R2R-CE |
| GPT-5.6-Sol | 1.7 | 2.2 | 3.0 | 1.4 | 0.163 |
| OmniBrain-0 (4B) | 0.65 | 1.2 | 0.71 | 0.13 | 0.042 |
| OmniBrain-0 (9B) | 0.38 | 0.86 | 0.85 | 0.11 | 0.044 |
Table 8: Average token consumption per task (millions of tokens, M).
Figure 9: Inference-time reminder-policy ablation. Success rates for execution without reminders and with the state-dependent reminder policy. The panels report Overall for LIBERO-Pro, LIBERO-Plus, RoboCasa365, and RoboTwin 2.0, and SR for R2R-CE.
Figure 10: Data-pipeline ablation on LIBERO-Pro. Success rates for removing reasoning reconstruction, removing recovery trajectories, and using the full data pipeline across task and swap conditions.
Figure 11: Experience-library ablation. Success rates with and without access to the Experience Library on LIBERO-Pro, RoboTwin 2.0, and RoboCasa365.
Figure 12: Examples of OmniBrain-0 on LIBERO-Pro. Top: turning on the stove and placing a moka pot on it. Bottom: placing a book in the back compartment of a caddy. Red annotations highlight the failed grasp.
Figure 13: Examples of OmniBrain-0 on RoboTwin 2.0. Top: arranging colored blocks in a row. Bottom: opening a microwave.
Figure 14: Examples of OmniBrain-0 on RoboCasa365. Top: stirring vegetables. Bottom: arranging a bread basket. Red annotations highlight a failed attempt in the vegetable task.
Figure 15: Examples of OmniBrain-0 on R2R-CE Val-Unseen. Top: navigating upstairs to a bathroom mirror. Bottom: navigating via a patio to a kitchen.
Figure 16: Real-world manipulation examples. Selected observations and structured decisions illustrate closed-loop perception, skill coordination, and execution feedback on a physical robot.
Figure 17: Reminder-guided manipulation. A wine-bottle placement example aligning visual observations (top), reminder-policy trigger states (middle), and OmniBrain-0’s tool invocations (bottom). For the task Put the wine bottle in the bowl , reminders are triggered at steps 2 and 3 before grasp and move_arm , respectively; the remaining displayed steps proceed without reminders. The sequence continues through bottle transport, release, and task termination.
Figure 18: Reminder-guided navigation. An illustrative bathroom-mirror navigation example aligning visual observations (top), reminder-policy trigger states (middle), and OmniBrain-0’s tool invocations (bottom). Reminders at steps 3 and 6 guide route selection and prevent premature termination, respectively.
Figure 8.1 : General Tools: Shared tools for environment observation, pixel-to-metric localization, text-guided object segmentation, execution termination, and mobile-base control.
Figure 8.2 : VLA-Specific Tools: Tools for end-effector motion, wrist orientation, gripper control, and multi-step manipulation or interaction through a learned VLA policy.
Figure 8.3 : VLN-Specific Tools: Tools for panoramic observation, instruction following, navigation-goal grounding and execution, point-directed navigation, and trajectory backtracking.
Figure 10.1 : Training data composition. (a) The VQA corpus contains 820,453 samples. (b) The initial agentic corpus contains 3,846 demonstrations. (c) RSI contributes 1,142 additional demonstrations. Percentages indicate each source’s share within the corresponding corpus.
Figure 11.1 : LIBERO-Pro task execution. Complete execution process for the task Pick the salad dressing and place it in the basket .
Figure 11.2 : RoboTwin 2.0 task execution. Complete execution process for placing the two specified bread loaves in the beige breadbasket.
Figure 11.3 : RoboCasa365 task execution. Complete execution process for placing a croissant and a jam jar together on the cutting board.
Figure 11.4 : R2R-CE task execution. Selected stages of a navigation task in the Val-Unseen split, illustrating landmark grounding, instruction following, obstacle avoidance, and endpoint verification.
Figure 11.5 : R2R-CE navigation comparisons. Each sequence compares InternVLA-N1 (left) and OmniBrain-0 (right) on the same Val-Unseen task. Cells pair front-view observations with top-down trajectories at the indicated environment steps. Blue annotations show OmniBrain-0’s decisions, and gray annotations describe InternVLA-N1’s behavior.
More in this topic
Embodied AI
Embodied AI studies agents that perceive, reason, and act through bodies in physical or simulated environments.
See all 255 papers