# Task B Fast B2-Piper Locomotion Training

## Goal

Train a stronger B2-Piper locomotion model for the EE-camera reach pipeline:

- Straight waypoint transit: target up to `2.0 m/s`, Piper fixed on the back.
- Target approach: target up to `1.0 m/s`, Piper arm randomized so the leg policy learns to tolerate gripper motion.

These are training tasks only. They do not modify official Task B scoring or object spawn logic.

## Added Gym Tasks

```text
ATEC-Isaac-Velocity-Flat-Unitree-B2Piper-FastStraight-v0
ATEC-Isaac-Velocity-Flat-Unitree-B2Piper-ApproachRandomArm-v0
```

The policy still observes and controls the 12 B2 leg joints. The Piper arm is part of the physics body:

- In `FastStraight`, the arm starts in the fixed backpack pose.
- In `ApproachRandomArm`, the arm joints are randomized at reset via a training-only event.

## Training Commands

Straight 2 m/s curriculum:

```bash
micromamba run -n atec2026 python scripts/rsl_rl/train.py \
  --headless \
  --task ATEC-Isaac-Velocity-Flat-Unitree-B2Piper-FastStraight-v0 \
  --num_envs 4096 \
  --max_iterations 8000
```

Approach 1 m/s random-arm curriculum:

```bash
micromamba run -n atec2026 python scripts/rsl_rl/train.py \
  --headless \
  --task ATEC-Isaac-Velocity-Flat-Unitree-B2Piper-ApproachRandomArm-v0 \
  --num_envs 4096 \
  --max_iterations 8000
```

## Validation Gates

1. Smoke-start each env with a small `--num_envs` and short iteration count.
2. Export TorchScript policies.
3. Test straight velocity tracking at commanded `2.0 m/s`.
4. Test approach velocity tracking at commanded `1.0 m/s` while the Piper arm pose changes.
5. Integrate the better approach policy into the Task B EE-lock controller only after it passes validation.