README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - robotics
  - vla
  - robosuite
  - imitation-learning
  - flock
dataset_info:
  features:
    - name: episode_index
      dtype: int32
    - name: step_index
      dtype: int32
    - name: task
      dtype: string
    - name: difficulty
      dtype: string
    - name: instruction
      dtype: string
    - name: image
      dtype: image
    - name: action
      sequence: float32
    - name: proprio
      sequence: float32
    - name: reward
      dtype: float32
    - name: done
      dtype: bool
  splits:
    - name: train
      num_examples: 25328

FLock Robotics VLA Training Dataset v2

Expert demonstrations for the FLock Robotics VLA competition task. All trajectories are successful.

Quick start

from datasets import load_dataset

ds = load_dataset("random-sequence/flock-robotics-vla-training-v2")
print(ds["train"][0])
# Keys: episode_index, step_index, task, difficulty, instruction,
#       image (PIL), action [7], proprio [25], reward, done

Dataset statistics

Task	Episodes	Difficulty	Steps
lift_cube	8	low	~1,100
pick_place_can	20	low	~3,800
pick_place_milk	19	low	~3,600
pick_place_bread	18	low	~3,300
pick_place_cereal	16	low	~2,900
stack_blocks	22	medium	~10,600
Total	103		25,328

All 103 episodes are 100% successful (success-filtered collection: each episode was retried until success).

Schema

Each row is one timestep:

Column	Type	Description
episode_index	int32	Trajectory ID (0–102)
step_index	int32	Timestep within episode
task	string	Task name (e.g. lift_cube)
difficulty	string	low or medium
instruction	string	Natural language task instruction
image	Image	224×224 RGB agent-view frame
action	float32[7]	OSC delta [Δx, Δy, Δz, Δroll, Δpitch, Δyaw, gripper] clipped to [-1, 1]
proprio	float32[25]	Robot state: joint pos/vel, EEF pose, gripper
reward	float32	Sparse reward (1.0 at success, 0 otherwise)
done	bool	Episode termination flag

Schema version: robotics_vla_sample_trajectory_v1

Environment

• Simulator: robosuite 1.5.2
• Robot: Panda 7-DOF
• Controller: BASIC composite (OSC_POSE arm + gripper)
• Camera: agentview 224×224

Raw zip

training_traces_v2.zip is also available for backward compatibility. It contains the same data as individual metadata.json + trajectory.npz pairs under trajectories/traj_NNN_<task>_<difficulty>/.

SHA256: e255a52aa0bb9fbf77a6d44ce368e41cb1d2dec184c2e7b6530e83088d772411