NAKA
CODE IS POLICY
Your robot doesn't need a trained policy. NAKA writes one.
VS PUBLISHED PAPER
2.2×
- 100% CUBE_LIFT
- 179 TESTS
- ZERO TRAINING
THE PROBLEM
The entire VLA field assumes you need 100K demonstrations, a billion-parameter policy, and hope it generalizes. When it fails on novel objects — and it always fails — you retrain. Months between "customer wants X" and "robot can do X."
The VLA community has been optimizing the wrong axis. The bottleneck isn't the model — it's the assumption that the model must contain the behavior.
This scales for Google. It does not scale for everyone else.
THE INVERSION
When ANIMA encounters a new task, NAKA doesn't search for a pre-trained policy. It writes one — in Python, in real time.
No training. No fine-tuning. No dataset collection. Code generation, visual feedback, and iteration.
from anima import azoth, ergon, haptos, prism
objects = azoth.detect(camera_image)
mug = [o for o in objects if o.label == "red mug"][0]
pose = ergon.estimate_pose(mug.crop, mug.depth)
grasp = compute_grasp(pose, gripper="parallel")
trajectory = ik_solve(grasp, current_joints)
robot.move(trajectory)
robot.close_gripper()
contact = haptos.verify_grasp()THE RESULTS THAT MATTER
CaP-BENCH COMPARISON — PASS@1, 50 EPISODES
| TASK | NAKA BEST | PAPER GEMINI 3 PRO | PAPER HUMAN EXPERT |
|---|---|---|---|
| cube_lift | 100% | 45% | 93% |
| cube_stack | 93% | 30% | 73% |
- 2.2×
- vs CaP-X paper on cube_lift
- 2.7×
- vs paper on cube_stack
- 100%
- cube_lift — exceeds human expert (93%)
- 93%
- cube_stack — exceeds human expert (73%)
- 187
- manipulation tasks available
- 179
- tests passing on real MuJoCo sim
Benchmarked against the CaP-X paper (NVIDIA + Berkeley + Stanford + CMU)
INSIDE THE ANIMA STACK
NAKA is the only module in the ANIMA stack that consumes every other module. It is the first citizen of a radically different way to build robots.
PERCEPTION AS A LIBRARY
ABYSSOS, AZOTH, ERGON, SOL, HAPTOS, PRISM — all accessible as library imports. Scene understanding on demand.
from anima import *VLA POLICIES AS API CALLS
MORPHEUS, PYGMALION, TITAN — callable tools inside NAKA's generated code, not controllers of the robot.
policy.infer()VISUAL DIFFERENCING FEEDBACK
Before/after frames drive the code rewrite loop. The robot observes failure and rewrites its own plan.
diff(frame_a, frame_b)SKILL LIBRARY AUTO-SYNTHESIS
Successful NAKA-generated code becomes a named, reusable skill that future tasks can reference.
skills.register()
PRODUCTION-READY
- More than 2× published results.
- Exceeds human expert baselines.
- Zero RL training.
- 179 tests passing on real MuJoCo sim.