cagataydev/neon-wholebody-nav-15k

---
license: mit
task_categories:
- robotics
tags:
- neon
- vla
- humanoid
- g1
- navigation
- whole-body
- loco-manipulation
size_categories:
- 10K<n<100K
---

# neon-wholebody-nav-15k

**Whole-body navigation + manipulation** — 15K episodes of 32-DoF joint trajectories combining locomotion and arm control.

## Description

Each episode has **two phases**:
1. **Navigation phase**: Walking/turning with sinusoidal gait (locomotion commands active, arms swing naturally)
2. **Manipulation phase**: Arms execute task while legs maintain stance (locomotion stops)

Smooth phase blending ensures realistic transitions from walking to reaching.

Tasks include:
- Walk to location AND pick up object
- Navigate through space AND interact
- Locomotion + bimanual coordination

Each episode contains:
- `language_instruction`: Natural language task with navigation + manipulation
- `actions`: JSON array of joint trajectories (T × 32: 29 body + 3 locomotion)
- `nav_phase_length`: Timesteps in navigation phase
- `manip_phase_length`: Timesteps in manipulation phase

## Part of Neon VLA

Training data for [Neon](https://github.com/cagataycali/neon) — open-source Vision-Language-Action model for humanoid whole-body control.

Total Neon dataset collection: **175K episodes across 9 datasets (~1.4GB)**