Most people ask "how does synthetic data for robot learning actually work?" so here's the real walkthrough, no fluff.

1. You call the SDK, describe a task

from dataforge import Env
env = Env.generate(task="pick_and_place", objects=["mug", "block"], randomize=True)

That single call procedurally builds a simulated scene — lighting, object placement, physics all randomized so your model doesn't overfit to one layout.

2. The environment streams sensor data

Every step of the simulated rollout emits:

• RGB-D frames

• LiDAR point clouds

• Proprioceptive joint/motor state

3. You get labeled trajectories out the other end

No manual annotation. Every frame is paired with ground-truth action + pose labels, ready to drop into an imitation learning or RL pipeline.

4. Sample dataset — try before you commit

We're dropping a free 500-episode "pickandplace" sample dataset in the community this week so you can benchmark it against your own real-world data before subscribing. Drop a comment if you want early access.

This is the whole point of DataForge Pro: skip the robot lab, skip the teleoperation rig, and get straight to training.

Everyone assumes you need a $50k teleoperation rig and a warehouse full of sensors to train a decent manipulation policy. You don't.

The biggest models in robotics right now (RT-X, Octo, OpenVLA) all lean heavily on simulated + synthetic data mixed with a smaller slice of real demonstrations. The sim-to-real gap has narrowed enough that domain randomization gets you most of the way there for a huge range of tasks — grasping, navigation, basic manipulation.

What that means practically if you're an indie ML engineer or hobbyist:

• You can generate thousands of labeled trajectories (RGB-D, LiDAR, proprioception) overnight instead of weeks of manual data collection.

• You can iterate on task/environment design instantly instead of resetting physical hardware between runs.

• You can afford to fail fast — bad runs cost compute, not broken servos.

The bottleneck isn't hardware anymore, it's whether your data pipeline can keep up with how fast you want to iterate. That's the exact problem I built DataForge Pro to solve — spin up a sim environment, generate labeled data, export to your training pipeline, repeat.

If you're tinkering with robot learning on nights/weekends, curious what a synthetic-data-first workflow looks like — happy to talk shop.