> ## Documentation Index
> Fetch the complete documentation index at: https://docs.innate.bot/llms.txt
> Use this file to discover all available pages before exploring further.
# Evaluate and iterate
> Test your trained policy, diagnose problems, and improve it.
Training a policy is rarely one-and-done. This page covers how to evaluate your skill, identify common failure modes, and iterate toward reliable performance.
## First test
After [deploying your trained skill](/training/deploy-trained-skill), run it from **Manual Control** with the same setup you used for recording.
Place the robot, objects, and lighting as close as possible to the conditions you recorded in. The first test should be easy for the policy — if it fails on its own training distribution, something is wrong.
Select the skill in Manual Control and tap play. Watch the full execution without intervening.
Did the robot complete the task? Where did it hesitate, overshoot, or fail? Mental notes are fine — you'll iterate fast.
## Common failure modes
| Symptom | Likely cause | Fix |
| --------------------------------------- | -------------------------------------------------------- | -------------------------------------------------------------------- |
| Robot doesn't move or barely moves | Too few episodes, or episodes have inconsistent starts | Record more episodes with consistent start poses |
| Arm overshoots the target | Jerky demonstrations or high variance in approach angles | Re-record smoother demonstrations; try a larger chunk size |
| Robot starts well but drifts | Not enough variation in demonstrations | Add more episodes with slight object position changes |
| Works on first run, fails on repeat | Object or robot position shifted | Record with more position variation; aim for 2–5 cm spread |
| Gripper doesn't close at the right time | Inconsistent grasp timing across episodes | Focus on consistent timing when closing the gripper |
| Robot ignores the object entirely | Lighting or background changed significantly | Record in the current conditions, or control lighting more carefully |
## How to improve a policy
### Add more data
The most reliable way to improve a policy. Add 20–30 episodes that specifically cover the failure case, sync, and retrain. You don't need to start from scratch — the new episodes are added to the existing dataset.
### Tune hyperparameters
If the behavior is qualitatively close but not quite right, adjust the run configuration — chunk size for the smoothness/reactivity tradeoff, max steps for dataset size, learning rate for training stability. [When to change the defaults](/training/train-act-policy#when-to-change-the-defaults) maps each symptom to the right knob.
### Improve demonstration quality
Review your recorded episodes. Look for:
* Episodes where you hesitated or corrected course excessively
* Episodes that are much longer or shorter than average
* Episodes where the start pose is significantly different
Replace low-quality episodes with clean ones, re-sync, and retrain.
## Scaling up
Once your policy works in the original setup, gradually introduce variation:
1. **Move the object** a few centimeters between runs
2. **Change the object** slightly (same cup in a different color)
3. **Adjust lighting** modestly
If the policy breaks, record 10–20 more episodes under the new conditions and retrain. Each round of data makes the policy more robust.
Invest in data variety and you'll spend less time debugging — see [how many episodes you need](/training/data-collection#tips-for-high-quality-data) for concrete numbers.
## Structured evaluation
A measured version of this workflow is taking shape on the web app's Profiling page: scored rollouts with failure tags, dedicated evaluation datasets, and auto-stop for learned skills. Still under development — not in a released OS build yet.