Coverage
v1 covers LeRobot-centric robot learning. 38% of the grid has evidence and the rest is unknown. This page is generated from the data at build time.
Last full review 2026-10-24dataset 1.0.0-seed
504grid cells (24 robots × 21 policies)
193with evidence (38.3%)
44%of filled cells have T1 or T2 evidence
218of the unknown cells are untested but have both sides in LeRobot (inferred, not evidence)
0stale (>180 days) · 0 version drift
32measured runs across 7 compute targets
Cells by state
| works | 27 | 5.4% |
| works with caveats | 67 | 13.3% |
| partial | 99 | 19.6% |
| not supported | 0 | 0.0% |
| planned | 0 | 0.0% |
| no verified evidence | 311 | 61.7% |
Filled cells by highest evidence tier
| T1 · upstream documented | 85 | Official docs, model card or release notes state it |
| T2 · maintainer tested | 0 | A maintainer ran it and recorded config + result |
| T3 · community reported | 108 | Public, reproducible third-party report |
| T4 · inferred | 0 | Derived from specifications, not observed |
Evidence age
1930–30 d
031–90 d
091–180 d
0181–365 d
0>365 d
Known-weak areas
- Tactile and force sensing is thin because the data barely exists: few public policies consume tactile input, and almost no runs are published.
- Deployment vs training. Many cells are evidenced only by a public checkpoint (training works); on-robot inference is rarely documented by a primary source. Those cells are partial, not works.
- SO-100 vs SO-101 attribution. Since LeRobot v0.6, both arms share one config class and new datasets record robot_type=so_follower, so a checkpoint trained on them cannot be attributed to either arm. Only datasets recorded with the older so100_follower / so101_follower strings count as evidence; some real SO usage is therefore not shown.
- Edge measurements come from few independent sources. Treat single-source Jetson numbers as indicative of that exact runtime, precision, camera count and power mode.
- Robots with no evidence yet: EarthRover Mini Plus.
- Compute targets with no measured runs: NVIDIA A100 40GB, NVIDIA A100 80GB, NVIDIA A10G (AWS), Apple Silicon 16 GB, Apple Silicon 32 GB, Apple Silicon 64 GB, CPU only (no accelerator), NVIDIA H200, NVIDIA Jetson Orin NX 16GB, NVIDIA L4, NVIDIA L40S, NVIDIA GeForce RTX 3060 12GB, NVIDIA GeForce RTX 5090, NVIDIA T4.
What this does not cover
| Autonomous driving stacks | Different problem, different evaluation culture, mature dedicated references exist. |
| Drones / UAV | Flight-control policies share almost no interface with manipulation policies. |
| Industrial PLC integration | Deterministic control, not learned policies. |
| Surgical robotics | Regulated; no public policy configurations to verify. |
| Soft robotics | No standard policy interface exists yet. |
| Closed commercial humanoids (Optimus, Figure, 1X NEO) | No public configuration to verify. See /closed-humanoids. |
| Benchmark success rates of any kind | This is a compatibility reference, not a leaderboard. Use LeRobot eval, RoboArena or vla-eval. |
| Policies only on LeRobot main (unreleased) | FineART-VLA, LingBot-VLA 2.0, Wall-OSS-0.5, DM05, LaWAM, FLUX 3 Action are tracked once released. |
| Moss v1 | It was never in a LeRobot release. The `moss` type existed only on main, from Oct 2024 until the Hardware API redesign of June 2025. There are no Moss datasets or checkpoints in the lerobot Hub org, and the maker recommends the SO-100 instead. |
| Reachy Mini | Reachy Mini is not a LeRobot robot. It ships its own SDK (reachy_mini Python package, JavaScript SDK) and a Hugging Face Spaces app ecosystem. It is absent from LeRobot v0.6.1's supported-hardware list and robots source tree, and its SDK does not depend on LeRobot. It is an expressive desktop head/body robot with no arm or gripper, so manipulation policies do not apply. |
| Hello Robot Stretch 3 | Native LeRobot support was removed in 0.4.0 (last release with it: 0.3.3). There is only one Stretch dataset in the lerobot Hub org (lerobot/cmu_stretch) and no official checkpoint, so there is no current LeRobot path to evaluate policies on it. |
Scope expands beyond LeRobot (ROS 2, openpi-native, Isaac Lab) only after grid coverage passes 70%, and that change will be announced here.