Robotic Reinforcement Learning Environment

1 Competition Outline

Background

Houstin Dynamics builds robotic picking systems for warehouse fulfillment centers. A core bottleneck is generalizing grasp policies to novel, previously unseen objects without extensive per-SKU tuning. This competition provides a simulated pick-and-place environment (built on an open-source physics simulator) where competitors train and submit a policy for grasping a diverse set of household and warehouse objects.

Target & Benchmark

Competitors train a reinforcement learning agent to control a simulated 6-DOF arm with a parallel-jaw gripper, picking objects from a cluttered bin and placing them in a target zone. Policies are evaluated on a held-out set of 50 object meshes not included in the provided training set. The benchmark to beat is HD's internal PPO baseline policy, which achieves a 71.4% grasp success rate on the held-out set.

Dataset

Download Dataset

Scoring Rules

• Policies are evaluated over 500 held-out episodes using 50 object meshes not seen during training.
• Score is the grasp success rate: the fraction of episodes where the target object is successfully placed in the goal zone within 200 simulation steps.
• Submitted policies must run inference within 100ms per step on the provided reference GPU to be considered valid.
• Competitors must submit trained model weights and the exact inference code used to produce actions; training code is optional but encouraged for tie-breaking review.
• Any policy found to exploit simulator physics bugs rather than performing genuine grasps (as judged by a human review pass) will be disqualified.

2 Prize Structure

The prize pool is split across the top-ranked leaderboard finishers. All winners listed below also receive a guaranteed first-round interview with Houston Dynamics.

3 About Houston Dynamics

Houston Dynamics designs and deploys robotic picking arms for third-party logistics providers and e-commerce fulfillment centers across North America. The company's research group focuses on sim-to-real transfer, reinforcement learning, and grasp generalization, with deployed systems currently operating in 12 fulfillment centers.

4 Roles Being Hired

This competition is sourcing internship candidates for:

• Robot Learning Research Engineer — develop reinforcement learning and imitation learning methods for manipulation, and drive sim-to-real transfer onto physical hardware.
• Simulation Infrastructure Engineer — build and scale the training environments, domain randomization pipelines, and evaluation infrastructure used to develop and validate grasping policies.

Both roles are based in Houston, TX for the summer of 2027.

5 Fine Print & Rules

• This is NOT a real competition. This is just a placeholder for future competitions that may be hosted on SearchSpace.
• Submitted model weights and inference code may be reviewed and re-run by HD to confirm reported results; discrepancies beyond normal simulation variance may result in disqualification.
• The provided simulation environment is a modified, research-only version of HD’s internal training stack and does not reflect current production performance.
• HD employees, contractors, and immediate family members are not eligible to compete.
• Cash prizes are subject to applicable tax withholding and reporting requirements.
• Guaranteed interviews require passing standard identity verification and, where applicable, work authorization confirmation.

6 Submission Instructions

Submit your trained model weights, inference code, and a short (1-page max) description of your training approach (algorithm, reward shaping, domain randomization, etc.) as a single zipped archive.

Submissions using pretrained foundation models are permitted as long as the final policy is fine-tuned or adapted using reinforcement learning on the provided environment.