TVRBench
Agents move through simulated indoor scenes to reproduce a target image's viewpoint. Success requires stopping at exactly the target pose on the simulator's discrete grid. The evaluation split has 500 tasks across 80 scenes.
Training and reinforcement-learning tasks use separate scene splits. Visual-action history and action-only memory are distinct evaluation settings. The paper also studies post-training methods.
- Made by
- Zhejiang University
- Released
- May 2026
- Kind
- Simulated robots
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- A simulator
- Items from
- Generated
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
-
Gemini 3.1 Pro #40 on the capability index
- Last model added
- April 2026
- Scores refreshed
- November 18th, 2026
In the authors’ words
Humans can reproduce the viewpoint specified by a target image through active head and body motion, yet spatial intelligence in foundation models has largely been studied as passive understanding of pre-collected observations. We introduce Target Viewpoint Reproduction (TVR) – an active task where an agent adjusts its viewpoint in a 3D environment until its observation matches a given target image – and TVRBench, an indoor-simulation benchmark spanning scene scale and target-view visual richness. TVR is far from solved: on the evaluation split, the strongest open-source and closed-source models reach only 7.8% and 12.0% success. Fine-grained analysis identifies two consistent bottlenecks: off-the-shelf models struggle with multi-turn visual history, and performance drops sharply when viewpoint reproduction requires body translation rather than in-place rotation, exposing a gap in mapping spatial discrepancies to embodied movement. To study reducing this gap, we build a unified TVR post-training framework covering expert-trajectory SFT, rationale-supervised CoT-SFT, offline Single-turn GRPO, and on-policy Multi-turn GRPO from live simulator rollouts. Visual-action SFT supplies the main gain, raising a 9B open-source model to 50.8% success; Multi-turn GRPO provides targeted multi-room refinement and reaches 51.4% overall, while CoT supervision and Single-turn GRPO degrade closed-loop performance. These results establish TVRBench as a testbed for measuring and training foundation models that actively perceive and act in 3D environments. Our code, data, and models are available at https://github.com/aim-uofa/TVRBench.
Gemini 3.1 Pro leads with 12% among models released by April 2026 Gemini 3.1 Pro leads with 12% among models released by April 2026
Best score per entry from the benchmark's paper, read on November 18th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in April 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1
Gemini 3.1 Pro action-only memory 12%
- 2 GPT-5 visual-action memory 8.00%
- 3
Qwen3.5-27B action-only memory 7.80%
- 4
Qwen3.6 27B action-only memory 7.00%
- 5 GPT-4o (snapshot unspecified) action-only memory 5.20%
- 6
Qwen 3.6 35B-A3B action-only memory 4.80%
- 7
Qwen3.5-35B-A3B action-only memory 3.80%
- 8
Qwen3.5-9B action-only memory 2.80%
Source: The TVRBench paper
- 1
Gemini 3.1 Pro action-only memory 12%
- 2 GPT-5 visual-action memory 8.00%
- 3
Qwen3.5-27B action-only memory 7.80%
- 4
Qwen3.6 27B action-only memory 7.00%
- 5
Qwen 3.6 35B-A3B action-only memory 4.80%
- 6
Qwen3.5-35B-A3B action-only memory 3.80%
- 7
Qwen3.5-9B action-only memory 2.80%
Source: The TVRBench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 12% | action-only memory | Feb 2026 | Official leaderboard |
| 2 | GPT-5 | 8.00% | visual-action memory | Aug 2025 | Official leaderboard |
| 3 | | 7.80% | action-only memory | Feb 2026 | Official leaderboard |
| 4 | | 7.00% | action-only memory | Apr 2026 | Official leaderboard |
| 5 | GPT-4o (snapshot unspecified) | 5.20% | action-only memory | Official leaderboard | |
| 6 | | 4.80% | action-only memory | Apr 2026 | Official leaderboard |
| 7 | | 3.80% | action-only memory | Feb 2026 | Official leaderboard |
| 8 | | 2.80% | action-only memory | Feb 2026 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 12% | action-only memory | Feb 2026 | Official leaderboard |
| 2 | GPT-5 | 8.00% | visual-action memory | Aug 2025 | Official leaderboard |
| 3 | | 7.80% | action-only memory | Feb 2026 | Official leaderboard |
| 4 | | 7.00% | action-only memory | Apr 2026 | Official leaderboard |
| 5 | | 4.80% | action-only memory | Apr 2026 | Official leaderboard |
| 6 | | 3.80% | action-only memory | Feb 2026 | Official leaderboard |
| 7 | | 2.80% | action-only memory | Feb 2026 | Official leaderboard |
Rankings by part
We read 4 parts from the benchmark’s paper, each ranking models on one part of the test, such as Single easy, Single hard and Multi easy. GPT-5 leads 3 of them, and Gemini 3.1 Pro leads the other one. Arrows show how far a model moved from its overall rank among the models we track.
Single easy
7 models we track have a score here.
- 1 GPT-5 ↑1 22.4%
- 2
Gemini 3.1 Pro ↓1 21.6%
- 3
Qwen3.5-27B 14.4%
- 4
Qwen3.6 27B 12.8%
- 5
Qwen 3.6 35B-A3B 9.60%
- 6
Qwen3.5-35B-A3B 8.00%
- 7
Qwen3.5-9B 6.40%
Single hard
7 models we track have a score here.
- 1
Gemini 3.1 Pro 21.6%
- 2
Qwen3.5-27B ↑1 13.6%
- 3
Qwen3.6 27B ↑1 9.60%
- 4 GPT-5 ↓2 8.00%
- 5
Qwen 3.6 35B-A3B 8.00%
- 6
Qwen3.5-35B-A3B 5.60%
- 7
Qwen3.5-9B 4.80%
Multi easy
7 models we track have a score here.
- 1 GPT-5 ↑1 3.20%
- 2
Qwen3.6 27B ↑2 1.60%
- 3
Gemini 3.1 Pro ↓2 0.80%
- 4
Qwen3.5-27B ↓1 0.80%
- 5
Qwen 3.6 35B-A3B 0.80%
- 6
Qwen3.5-35B-A3B 0%
- 7
Qwen3.5-9B 0%
Multi hard
7 models we track have a score here.
- 1 GPT-5 ↑1 5.60%
- 2
Gemini 3.1 Pro ↓1 4.00%
- 3
Qwen3.6 27B ↑1 4.00%
- 4
Qwen3.5-27B ↓1 3.20%
- 5
Qwen3.5-35B-A3B ↑1 1.60%
- 6
Qwen 3.6 35B-A3B ↓1 0.80%
- 7
Qwen3.5-9B 0%
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The TVRBench paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| action-only memory | 12% | ||||
| GPT-5 | visual-action memory | 8.00% | |||
| GPT-5 | action-only memory | 8.00% | |||
| action-only memory | 7.80% | ||||
| action-only memory | 7.00% | ||||
| action-only memory | 4.80% | ||||
| action-only memory | 3.80% | ||||
| visual-action memory | 3.20% | ||||
| visual-action memory | 3.20% | ||||
| action-only memory | 2.80% | ||||
| visual-action memory | 0.80% | ||||
| visual-action memory | 0.20% | ||||
| visual-action memory | 0% |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The TVRBench paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| action-only memory | 12% | ||||
| GPT-5 | visual-action memory | 8.00% | |||
| GPT-5 | action-only memory | 8.00% | |||
| action-only memory | 7.80% | ||||
| action-only memory | 7.00% | ||||
| action-only memory | 4.80% | ||||
| action-only memory | 3.80% | ||||
| visual-action memory | 3.20% | ||||
| visual-action memory | 3.20% | ||||
| action-only memory | 2.80% | ||||
| visual-action memory | 0.80% | ||||
| visual-action memory | 0.20% | ||||
| visual-action memory | 0% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The TVRBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official TVRBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Robotics
- ERQA In use 400 multiple-choice visual questions from robot and egocentric footage on spatial, trajectory, action and state reasoning. Accuracy.
Gemini Robotics ER 2 78.5%
- SimplerEnv In use Real-to-sim replicas of Google Robot and WidowX/Bridge setups. Policies scored by success rate on pick, move, drawer and Bridge tasks. Praxis-1 80%
- RoboLab In use Zero-shot DROID Franka policies in Isaac Lab on 120 tabletop tasks at three instruction levels, 10 episodes each. Success rate. Praxis-1 46.2%
- RoboDojo In use Generalist manipulation policies run 42 Isaac Sim tasks and 18 remote real-robot tasks. Scored by task score and success rate. GPT-6 Astra 27.6%
- Progress classification (Google DeepMind) In use Models assign each frame of a robot video to one of five task-progress bands from 0 to 100 percent. Scored by accuracy.
Gemini Robotics ER 2 57.4%
- Human proximity (Google DeepMind robotics) In use Models judge from robot camera views whether a human is within 1 meter, a spatial awareness check for safe robot behavior. Scored by accuracy.
Gemini Robotics ER 2 93%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 7 months ago, in April 2026, for Qwen3.6 27B. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (8 entries).
- Pass Models scored 7 models we track, plus 1 other entry on its leaderboard.
- Warning Newest score The newest came 7 months ago, in April 2026, for Qwen3.6 27B.
- Note Frontier models None of the top 10 has a score. The strongest tested is Gemini 3.1 Pro (#40).
- Pass Headroom The best is 12%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 13 rows for models we track and 1 other entry at the last read.
- Pass Last read November 18th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.