Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -104,6 +104,14 @@ a stable cached latent appears to act as temporal regularization.
|
|
| 104 |
| **TinyVLA (this)** | **0.64B (140M online)** | **86% dual-rate** | **84.0%** |
|
| 105 |
| OpenVLA | 7B | 84.7% | 76.5% |
|
| 106 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
Honesty notes: our n=100 (±8pp) vs their 500; single suite; eval covers init states 0–9
|
| 108 |
per task (demos of same tasks/init states are in training data — this is the standard
|
| 109 |
LIBERO protocol used by all listed models). Multi-suite evals (object/goal/long) and an
|
|
|
|
| 104 |
| **TinyVLA (this)** | **0.64B (140M online)** | **86% dual-rate** | **84.0%** |
|
| 105 |
| OpenVLA | 7B | 84.7% | 76.5% |
|
| 106 |
|
| 107 |
+
**MolmoAct2 context** (from their cards/repo, no numeric tables published outside the
|
| 108 |
+
LIBERO-LeRobot card): 5B params on Molmo2-ER backbone + flow expert; claims to beat π0.5;
|
| 109 |
+
Molmo2-ER reportedly surpasses GPT-5 / Gemini-Robotics-ER-1.5 on 13 embodied-reasoning
|
| 110 |
+
benchmarks; #1 VLA on the MolmoSpace leaderboard. Inference needs ~26GB fp32 (16GB bf16),
|
| 111 |
+
cuda-graphs for ~2× speedup — H100-class serving vs our Raspberry-Pi-class 140M fast path.
|
| 112 |
+
Notably they ship **MolmoAct2-SO100_101** (absolute joint control) — a 5B competitor on our
|
| 113 |
+
target embodiment; a head-to-head offline eval on SO101 held-out data is an obvious next test.
|
| 114 |
+
|
| 115 |
Honesty notes: our n=100 (±8pp) vs their 500; single suite; eval covers init states 0–9
|
| 116 |
per task (demos of same tasks/init states are in training data — this is the standard
|
| 117 |
LIBERO protocol used by all listed models). Multi-suite evals (object/goal/long) and an
|