EmbodiedLab - T1 Process: leaderboard
EmbodiedLab (v0.1.0, arXiv 2610.35385) tests whether an agent can run a physical science experiment to find a quantity it cannot know in advance, such as a spring's stiffness or an acid's concentration. The agent gets an objective and the apparatus but no method, and writes Python that drives a simulated Franka arm in NVIDIA Isaac Sim and reads instruments through three cameras; the hidden parameters are redrawn for every instance. A run succeeds only if the submitted value is within the instrument's tolerance, rests on readings recorded during the run, and the bench is left as required. Each model runs in its developer's coding agent at that agent's middle reasoning setting, on 30 experiments at two interface levels: raw control (T0), and with perception and motion-planning tools (T1). Unlike manipulation benchmarks such as LIBERO, where the target state is given, the robot's actions here are only the means to a measurement. This board is the T1 process score, from 0 to 100, with perception and motion-planning tools.
Metric: T1 (perception and motion planning) process score (0-100). Source: hexafuture.ai. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Medium) | 73.7 | #5 (Claude Opus 5.5) |
| 2 | GPT-6 Astra (Medium) | 61 | #6 (GPT-6 Astra) |
| 3 | GPT-6.1 Sol (Medium) | 51.7 | #10 (GPT-6.1 Sol) |
| 4 | Kimi K3 (High) | 39.7 | #53 (Kimi K3) |
| 5 | DeepSeek V4.1 Flash (High) | 36.7 | #58 (DeepSeek V4.1 Flash) |
| 6 | GLM-5.3 Flash (High) | 20.7 | #119 (GLM-5.3 Flash) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=embodiedlab-t1-process · How It Works · Data refreshed daily, snapshot 2026-11-12.