04 · Benchmarks
vs Lightweight VLAs
Kinetic v0.3.0 · Generated 2026-07-16 · Windows AMD64 · Python 3.11.9
Methodology
Kinetic is a modular VLA stack (world model + optional small LLM + policy). Monolithic VLAs (OpenVLA, SmolVLA, …) differ in training and evaluation protocols. This report compares systems efficiency (latency, size class, sim success) and cites published numbers for peers, it is not a LIBERO leaderboard re-run.
Measured locally
- Control-loop stage latencies (p50/p95, ms)
- Effective control Hz from step p50
- Pick-bottle success rate over N trials (max_steps=400)
- Process RSS memory (approx, MB)
Literature-only
- Parameter counts and published Hz/VRAM for other VLAs
- Cross-paper success rates (LIBERO / real-robot), not same tasks as Kinetic sim
Task: Pick up the bottle from the table (Kinetic mock / MuJoCo sim)
Kinetic profiles
273.3Hz · Kinetic-mock-rule-heuristic
286.3Hz · Kinetic-mock-rule-mlp
202.5Hz · Kinetic-mujoco-rule-heuristic
| Profile | Lang | Policy | Robot | p50 ms | p95 ms | Hz | Success | Trials | RSS MB |
|---|---|---|---|---|---|---|---|---|---|
Kinetic-mock-rule-heuristic | rule | heuristic | mock_sim | 3.66 | 4.33 | 273.3 | 100% | 10 | 66.6 |
Kinetic-mock-rule-mlp | rule | mlp | mock_sim | 3.49 | 3.90 | 286.3 | 100% | 10 | 66.5 |
Kinetic-mujoco-rule-heuristic | rule | heuristic | mujoco | 4.94 | 5.98 | 202.5 | 100% | 5 | 157.6 |
Stage breakdown
Fastest profile: Kinetic-mock-rule-mlp
| Stage | p50 (ms) | p95 (ms) | mean (ms) |
|---|---|---|---|
vision | 1.148 | 1.311 | 1.173 |
scene | 0.123 | 0.159 | 0.123 |
policy | 0.003 | 0.078 | 0.036 |
control | 0.036 | 0.042 | 0.035 |
step_total | 3.492 | 3.899 | 3.518 |
reason | 0.018 | 0.018 | 0.018 |
Literature baselines
| Model | Params | Hz | VRAM | Architecture | Notes |
|---|---|---|---|---|---|
| OpenVLA-7B | 7000M | 5.0 | 16.0 | Monolithic VLM→action tokens | Strong open baseline; heavy for edge. |
| MiniVLA-1B | 1000M | 12.5 | - | Compact OpenVLA-style | ~2.5× faster than OpenVLA on L40s. |
| TinyVLA | - | - | - | Compact VLA family | Faster inference & better data efficiency than OpenVLA (paper). |
| Octo-Base | 93M | - | - | Transformer + diffusion (OXE) | Lightweight generalist. |
| Octo-Small | 27M | - | - | Transformer + diffusion (OXE) | Smallest common Octo config. |
| SmolVLA-450M | 450M | - | - | Compact HF VLA | Consumer hardware; async option. |
| OpenVLA-OFT | 7000M | - | 15.9 | OpenVLA + OFT | ~25–50× action gen speedup claims vs base. |
| Diffusion Policy | - | - | - | Visuomotor diffusion (IL) | Not full VLA; strong manipulation baseline. |
| ACT | - | - | - | Action chunking transformer (IL) | Bimanual IL; language usually separate. |
Head-to-head (systems efficiency)
| Dimension | Lightweight VLAs (lit.) | Kinetic (measured) |
|---|---|---|
| Parameters (control path) | 27M–7B monolithic | No monolithic VLA weights required for rule+heuristic path; optional BitNet ~0.7B only on replan events |
| Control rate | OpenVLA ~5 Hz; MiniVLA ~12.5 Hz (L40s) | up to 286 Hz (step p50, this host) |
| Sim pick success | n/a (different protocols) | 100% on Kinetic bottle task |
| LLM every control step | Usually yes (token decode) | No, reasoner event-only (rule / BitNet on replan) |
| Edge deploy story | Quantize large VLM | Modular: INT-ready slots + tiny control loop |
Takeaways
- Kinetic is not a drop-in OpenVLA clone, it is a modular stack optimized for low control-loop latency and optional language reasoning.
- On this host, the mock rule+heuristic path achieves multi-hundred Hz effective control with high sim pick success, orders of magnitude faster than 7B token-decoding VLAs.
- Fair skill comparison with OpenVLA/SmolVLA requires shared benchmarks (LIBERO/real robot) and trained policies; use Kinetic's MLP/diffusion slots + OXE-style data for that track.
- Use literature Hz only as order-of-magnitude; always re-measure on your hardware.
Numbers on this page are literature comparisons and local systems measurements for Kinetic, not product marketing claims.