Embodied models, three kinds, tested by us.
We sort embodied models by what they do and score each kind separately, on tests we run ourselves. Results from other public leaderboards are shown further down for reference; we never average them into our scores, because each board tests different models on different tasks.
Embodied reasoning
Phase 1The robot's eyes and planner: vision-language models that look at a scene, work out where things are and what can be done, and plan the steps.
Can it understand the scene and plan the task?
- Examples
- Gemini Robotics-ER, Cosmos Reason, RoboBrain
- How we test
- Private test set, run through public APIs and open weights
Robot policy
Phase 2Models that output motor commands directly (VLA): given camera images and an instruction, they move the arm.
Can it actually do the task on a real arm?
- Examples
- π0.5, GR00T, DM0, LingBot-VLA
- How we test
- Same task set on standard arms and in simulation, with partner labs
World model
Phase 2Models that predict what happens next when the robot acts, used for training and planning.
Are its predictions physically right?
- Examples
- Cosmos 3, V-JEPA 2, LTX-2.5, Odyssey-2
- How we test
- Open weights and APIs; Genie 3 has no public access yet
Embodied Reasoning Index · v0
Candidate models and run status
| Model | Developer | Access | Score | Status |
|---|---|---|---|---|
| Gemini Robotics-ER 1.6 (preview) | Google DeepMind | Gemini API · $1 in / $5 out per M tokens | — | Not yet run |
| Cosmos Reason2 (2B / 8B / 32B) | NVIDIA | Open weights · NVIDIA Open Model License | — | Not yet run |
| RoboBrain 2.5 (4B / 8B) | BAAI | Open weights · Apache-2.0 | — | Not yet run |
| Frontier general VLMs (baseline) | Google, OpenAI, Anthropic, Alibaba | API · latest version at run time | — | Not yet run |
Scores appear only after a full run with confidence intervals · Test items stay private so they cannot be trained on
Want your model on our bench, or a benchmark we don't cover yet? We run custom evaluations.Talk to us →Coverage matrix
85% of robot policies have a score on only one benchmark
Of 1,224 policies with published simulation scores, 1,047 appear on one benchmark and 23 on three or more; none appears on all four of LIBERO, SimplerEnv, CALVIN and RoboCasa. The same policy's LIBERO score also varies by paper: π0 ranges from 37.4% to 96.8% across 73 reports. A weighted mix of these boards would mostly measure how the gaps are filled.
| Model | RoboChallenge | RoboArena | BEHAVIOR | RoboLab | ArmnetBench | SimFoundry | FlagEval-Robo | LIBERO* | # |
|---|---|---|---|---|---|---|---|---|---|
| π0.5 | 8 | ||||||||
| π0 | 7 | ||||||||
| GR00T N1.6 | 5 | ||||||||
| GR00T N1.7 | 4 | ||||||||
| DreamZero | 5 | ||||||||
| OpenVLA-OFT | 3 | ||||||||
| SmolVLA | 3 | ||||||||
| MolmoAct2 | 4 | ||||||||
| X-VLA | 2 | ||||||||
| LingBot-VLA | 1 | ||||||||
| DM0 | 1 | ||||||||
| Spirit v1.5 / v1.6 | 2 | ||||||||
| Cosmos3 Policy | 2 |
Listed not listed not disclosed by the board
A dot means a published score on that board. *LIBERO scores are self-reported in papers. RoboArena lists Spirit v1.6, RoboChallenge v1 listed v1.5. Sources: RoboChallenge and RoboArena live boards (2026-10-28), FlagEval-Robo (2026-08), VLA Leaderboard dataset (2026-10-01), AllenAI vla-eval (2026-08-10).
Public leaderboards
RoboChallenge Table30
Real robotRun by Dexmal + Hugging FaceInference: submitter-hosted57 entries (V2)
Real-robot tabletop board. V2 (2026) adds 18 bimanual tasks, a DOS-W1 platform, zero-shot tests and completion-time scoring, and requires one generalist model.
Table30 V2 success rate, % (official ranking order); top score 56.80
Read with care. Inference runs on the submitter's servers, so the platform cannot confirm which model ran. The operator, Dexmal, also builds models: VLA-DM0.5 is fine-tuned from Dexmal's DM0.5, and the RoboChallenge baseline is π0.5. Only 4 of 57 entries pass 20% success.
Snapshot 2026-10-28 · Source: robochallenge.cn
BEHAVIOR Challenge 2025
SimulationRun by Stanford SVLInference: host-evaluated18 teams
50 household tasks in BEHAVIOR-1K on OmniGibson. Q-score gives partial credit for completed sub-goals.
Held-out Q-score
Read with care. Ranks competition teams, not released models, so entries rarely match other boards.
Snapshot 2025 final · Source: BEHAVIOR leaderboard
RoboLab-120
SimulationRun by NVIDIA SRLInference: host-run9
Simulation board on Isaac Sim 5.0: 120 manipulation tasks across visual, procedural and relational axes.
Success rate, %
Read with care. Run by NVIDIA; NVIDIA's own policies rank first.
Snapshot 2026-08-12 · Source: RoboLab data
FlagEval EmbodiedVerse-Open
Benchmark paperRun by BAAIInference: host-run24 VLMs
Embodied reasoning for vision-language models: the average of 13 spatial and planning benchmarks (Where2Place, RoboSpatial, VSI-Bench, ERQA, EgoPlan and others).
Average score across 13 benchmarks
Read with care. Data dates from November 2025, so current models such as Gemini Robotics-ER 1.6 are missing. Embodied-specific Cosmos-Reason1-7B scores 42.11, below general models. This is the closest public reference for our Embodied Reasoning Index.
Snapshot 2025-11 · Source: flagevalrobo.baai.ac.cn
SimFoundry
Real robotPaper authorsInference: author-run5
Real-to-sim policy evaluation; real-robot success reported alongside simulation.
Real-world avg. success, %
Read with care. One paper's runs, one snapshot.
Snapshot 2026-09-18 · Source: arXiv 2606.28276
ArmnetBench v0.1
Benchmark paperPaper authorsInference: author-run7
Arm manipulation benchmark; strict success counts only full completion.
Strict success, %
Read with care. One paper's runs, one snapshot.
Snapshot 2026-08-29 · Source: arXiv 2607.24481
FlagEval-Robo
Sim + realRun by BAAIInference: host-run, equal post-training13 open models (sim) · 6 (real)
Sim board averages LIBERO, VLA-Arena, ManiSkill and RoboCasa; real board uses ATOM-Bench atomic skills and composed tasks.
Sim board: mean success of LIBERO, ManiSkill, VLA-Arena, RoboCasa, % (DreamZero unranked, two benchmarks only)
Real-robot board (ATOM-Bench): mean of four PSR values, %, Franka single-arm + Cobot Magic dual-arm
Read with care. Open models only, each post-trained on the same evaluation data. LIBERO is saturated (most models 94–99%), so the spread comes from the other three. ACoT-VLA is AgiBot's model.
Snapshot 2026-08 · Source: flagevalrobo.baai.ac.cn
RoboArena
Real robot7 universities, DROID platformInference: submitter-hosted24 policies
Evaluators pick their own tasks and run double-blind pairwise comparisons; rankings come from preferences.
Rating from pairwise preferences (hover for ± std and number of evaluations); policies with fewer than 20 evaluations hidden
Read with care. Ratings are relative and noisy at low counts: the top entry, Spirit v1.6, has 25 evaluations and ±106, so its lead over DreamZero (190 evaluations) is not significant.
Snapshot 2026-10-28 · Source: robo-arena.github.io
LIBERO
SimulationSelf-reported in papersInference: self-reported1,607 reported results
Four simulated tabletop suites (Spatial, Object, Goal, Long).
Four-suite avg. success, % (self-reported)
Read with care. Top scores sit at 97–99%, so it no longer separates strong models; training data and protocols differ by paper.
Snapshot 2026-09-23 · Source: AllenAI vla-eval