Datasets:

ahive
/

ledrone_pybullet_hover

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - ledrone
  - uav
  - drone
  - sim
  - pybullet
  - hover
  - imitation-learning
  - goal-conditioned
  - state-based
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files: data/*/*.parquet

ledrone_pybullet_hover

A goal-conditioned, state-based imitation-learning dataset of a quadrotor performing the hover task (reach a fixed target position and station-keep), generated in PyBullet rigid-body physics with a privileged analytic expert. Each frame pairs the drone's 13-dim proprioceptive state and a relative goal with the expert's velocity + yaw-rate command — ready to train goal-conditioned control policies (Diffusion Policy, ACT, …). A first-person-view (FPV) camera is shipped as an optional auxiliary stream (the task is solvable from state alone).

Built with LeRobot (format v3.0) as part of the ledrone project.

At a glance


Task	`hover` — reach a fixed target position and station-keep
Type	Goal-conditioned, state-based (FPV image is an optional auxiliary)
Episodes	300
Frames	60,000 (200/episode, 6.7s at 30 Hz)
Robot type	`ledrone` (quadrotor)
Physics	PyBullet rigid-body dynamics via gym-pybullet-drones
Expert	Analytic PD + reference-velocity feed-forward + course-to-goal yaw, slew-limited startup (privileged, state-only)
Camera	FPV 96×96 RGB (AV1) — auxiliary; scene has a goal marker, landmarks, textured ground
Action	NED velocity + yaw rate
Format	LeRobot `v3.0`

How it was generated

A privileged analytic state-only expert (PD + reference-velocity feed-forward) flies the hover reference in the LeDrone env on the PyBullet backend.

DART-style coverage (Laskey et al. 2017): the drone is driven by a noised command (σ = 0.15 m/s on the velocity channels) but each frame is labelled with the clean expert command — so the policy sees off-distribution states and the optimal correction for them, mitigating behavior-cloning covariate shift.
Slew-rate-limited setpoint ramps the command from rest, taming the low-level controller's start-of-episode transient (early-episode tilt/rate spikes).
Controlled heading (course-to-goal yaw); randomized start + goal per episode; the goal is stored relative to current position (origin-invariant).
Convergence gate: an episode is accepted only once the expert reaches / tracks the goal (else the start/goal is resampled).

Reproduce (needs the [pybullet] extra):

python tools/generate_expert_dataset.py --task hover --backend pybullet \
    --repo-id ahive/ledrone_pybullet_hover --episodes 300 --push-to-hub --public

Features

Key	Dtype	Shape	Description
`observation.state`	float32	(13,)	`pos_n/e/d`, `vel_n/e/d`, attitude quat `q_w/x/y/z`, FRD body rates `rate_x/y/z`
`observation.environment_state`	float32	(3,)	Goal relative to current position: `err_n`, `err_e`, `err_d`
`action`	float32	(4,)	Command: `vx`, `vy`, `vz`, `yaw_rate`
`observation.images.fpv`	video	(96, 96, 3)	Forward-facing FPV camera, AV1 (auxiliary)

State is the SOTA quadrotor policy observation (relative goal + velocity + orientation + body rates); battery is intentionally excluded (handled by domain randomization, not observed — Hwangbo 2017, Molchanov 2019, Kaufmann ICRA 2022 / Nature 2023). Attitude is a unit quaternion; map to a continuous 6D rep (Zhou et al., CVPR 2019) if desired. action is a NED velocity + yaw rate clipped to the envelope (≤ 2 m/s horizontal, ≤ 1 m/s climb, ≤ 45 °/s yaw).

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("ahive/ledrone_pybullet_hover")
frame = ds[0]  # observation.state (13,), observation.environment_state (3,), action (4,)

Intended use & limitations

Intended use: goal-conditioned, state-based imitation for quadrotor hover (state + relative goal → velocity + yaw-rate). 300 episodes, all in train.
State-based, not visuomotor. The task is solvable from observation.state + observation.environment_state; the FPV goal marker is in only ~7% of frames. The image is an optional auxiliary input, not the primary task signal.
DART coverage, but no hard failures. The injected noise gives off-distribution recovery labels, but the data still contains no crash / large-perturbation / failure demonstrations.
Simulation only (PyBullet / gym-pybullet-drones, Crazyflie 2.x). The FPV scene is a synthetic benchmark scene, not photorealistic / sim-to-real imagery.

Data analysis

Analysis across all three ledrone tasks, computed from this DART-regenerated data.

Signature trajectories — takeoff climbs from the ground, hover converges & station-keeps, circle flies full loops (>= 360 deg).

Convergence & attitude — every task drives the goal error to near-zero (circle keeps a small steady tracking lag from following a moving target); body tilt stays gentle (median ~10 deg, 95% <= 35 deg).

Command / action space — takeoff & hover station-keep at ~0.1 m/s while circle orbits at ~1.3 m/s; the climb command is bimodal for takeoff; yaw-rate is centered on the goal bearing.

License & citation

Apache-2.0. Please cite the ledrone project.