# Pose-transfer adapter

The repository includes two complementary mechanisms for teaching semantic
motion while respecting the Dropbear articulation.

## Playback constraint projection

`dropbear_walk/pose_guidance.py` defines named stand and crouch poses and a
test sequence:

```text
stand -> walk with arm swing -> crouch -> stand up -> walk
```

`ConstrainedPoseGuide` maps named semantic targets into the 14-dimensional
policy action order, clamps targets to softened mechanical joint limits, and
rate-limits guided joints below their configured actuator velocity limits.
During walking, the policy retains leg ownership and the adapter guides only
reciprocal shoulder swing.

This is useful for visualization, mechanical-range discovery, and generating
curriculum references. It is not a substitute for a learned balance policy.

## Learned residual pose control

The environment also supports pose-reference observations and reward terms for
joint pose, COM height, and height-trajectory velocity. The training wrapper
can:

- append pose/COM feedback channels to older checkpoints;
- zero-initialize added network columns to preserve inherited behavior;
- train a policy residual around a semantic pose;
- bias pelvis pitch, knees, and ankle joints by name;
- stagger pose phases across vectorized environments;
- evaluate stand/crouch phases and report joint/height errors.

Relevant flags begin with `--pose-*`, `--com-*`, `--adapt-pose-checkpoint`,
and `--calibrate-*-normalizer`.

## What v0.1.0 does and does not claim

The included `model_8745.pt` is the validated upright planar locomotion
checkpoint. The pose-transfer stack is included and has been exercised in
simulation, but crouched locomotion and floor-to-stand recovery are not
release-qualified policies in v0.1.0.

Recommended progression:

1. hold zero velocity while lowering COM height smoothly;
2. recover to standing at zero velocity;
3. add low-speed forward/back/left/right commands at fixed crouch height;
4. randomize target height within mechanically reachable bounds;
5. train ground recovery separately with staged initial poses, contact-aware
   rewards, and strict actuator/impact limits;
6. combine skills only after each policy passes deterministic transition
   gates.