RWKV-GroupKAN Flow-Matching Vision-Language-Action Model
A novel <400M parameter VLA for few-shot robotic manipulation, targeting 90%+ success on MetaWorld MT-50 with only 10 expert demonstrations per task.
Vision (SigLIP-base, frozen) + Language (SmolLM-135M, frozen)
↓ Cross-Attention Fusion
Condition Encoder (pool + proprio + time embed)
↓
RWKV-KAN U-Net (3-stage encoder-decoder, ~33M params)
├── RWKV: Linear-time temporal + channel mixing (bidirectional scan)
└── GroupKAN: B-spline KAN with G=4 groups + Channel Affinity Modulation
↓ velocity field v_θ(a_t, t, condition)
Multi-Segment Consistency Flow Matching (K=2, one-step inference)
└── Action Consistency Regularization (expert anchoring)
↓
Action output (horizon=4, 3D delta + gripper)
First VLA to combine RWKV + GroupKAN + Consistency Flow Matching. Based on the "KAN-We-Flow" paper (arXiv:2602.01115v2, Feb 2026) which achieves SOTA on MetaWorld/Adroit/DexArt with only 33.6M parameters and ~10ms inference.
# Install
cd RWKV
pip install -e "."
# Smoke test (synthetic data, no GPU needed)
python -m kanflow_vla.train --smoke-test --batch-size 4 --mixed-precision fp32
# Full training on MT-50
python -m kanflow_vla.train \
--config kanflow_vla/configs/metaworld.yaml \
--batch-size 128 \
--epochs 3000 \
--wandb
# Evaluate
python -m kanflow_vla.eval_metaworld \
--checkpoint checkpoints/kanflow_vla/best.pt \
--num-rollouts 10| Difficulty | Target SR1 | Paper SR1 |
|---|---|---|
| Easy | ≥92% | 92.0±1.0% |
| Medium | ≥85% | ~85% |
| Hard | ≥78% | ~78% |
| Very Hard | ≥71% | 71.3±1.0% |
| Component | Params | Trainable |
|---|---|---|
| SigLIP-base (frozen) | ~86M | 0 |
| SmolLM-135M (frozen) | ~24M | 0 |
| Fusion (2-layer xattn) | ~5M | ~5M |
| Proprio MLP | ~33K | ~33K |
| RWKV-KAN UNet | ~33M | ~33M |
| Total | ~148M | ~38M |
- KAN-We-Flow — RWKV-KAN + Flow Matching for robotics
- SmolVLA — Compact VLA architecture
- RWKV — Linear-time sequence modeling
- KAN — Kolmogorov-Arnold Networks
Research use only. See individual component licenses.