About
Post-training is a crucial stage in robotics foundation model research: it enables robots to go beyond large-scale pretraining on expert demonstrations and further improve performance through their own interaction with the environment. By collecting interaction rollouts, distinguishing good and bad behaviors, and iteratively optimizing policies, post-training can substantially enhance precision, failure recovery, and generalization—bringing robotics foundation models closer to reliable real-world deployment.
This challenge and workshop will focus on the foundational questions and key challenges in the post-training stage of robotics foundation models. The event consists of two tightly coupled components: (1) a pre-workshop challenge and (2) an in-person workshop at RSS. The workshop program will include four parts: the pre-workshop challenge, invited speaker talks, team presentations, and a panel discussion.
We expect the workshop to (i) establish shared problem definitions and evaluation protocols for robotics foundation model post-training, (ii) surface empirical lessons and failure modes from real-robot rollouts, and (iii) catalyze reusable resources (e.g., benchmarking protocol, reports, and baselines) that enable more reproducible and scalable research in this area.
Challenge
The challenge is structured in two rounds. In Phase 1, all teams are evaluated on a standardized real-robot bimanual manipulation benchmark using a shared dataset. In Phase 2, the top teams are selected and supported in iterative rollout collection to further improve their policies via post-training.
Phase 1 — Evaluation Tracks
Phase 1 evaluates submitted policies on 3–5 standardized real-robot tasks across two ranking tracks. Teams perform offline training on the released dataset (expert data + baseline success and failure rollouts) and submit policies for benchmark evaluation.
Single-Task Track
Policies are evaluated and ranked on each task independently. Best for specialists optimizing for individual task performance.
Multi-Task Track
A single policy is evaluated jointly across all tasks. Best for generalist policies that share representations across skills.
Task Samples & Scoring
Phase 1 includes three real-robot bimanual manipulation tasks. Select a task below to view a sample rollout video and its scoring criteria.
Scoring Criteria · insert-mouse-battery
Place the mouse in the center of the table with the battery slot facing right.+0.4
Insert the battery in the correct polarity, but not fully inserted.+0.4
Battery is fully inserted.+0.2
Steps are cumulative — completing the final step yields a total score of 1.0.
Left wrist
Top
Right wrist
Scoring Criteria · tower-of-hanoi-game
Successfully place the small ring onto the opposite peg.+0.3
Successfully place the large ring onto the middle peg.+0.3
Successfully place the small ring onto the middle peg and stack it on top of the large ring.+0.4
As long as the final configuration has the large ring on the bottom and the small ring on top, the score will be counted as 1.0.
Left wrist
Top
Right wrist
Scoring Criteria · seal-water-bottle-cap
Insert the straw of the bottle cap into the bottle and place the cap onto the bottle.+0.5
Tighten the cap — score scales with how firmly it is fastened.+0 → +0.5final 0.5 – 1.0
The tightening score is graduated based on the number of turns the robot makes. A final score of 1.0 indicates the cap is fully tightened.
Left wrist
Top
Right wrist
Dataset
For Phase 1, we release a standardized real-robot bimanual dataset designed for offline training. The dataset contains four complementary components:
Standardized Real-Robot Bimanual Dataset
Expert Data
High-quality human teleoperation demonstrations on the benchmark tasks.
Baseline Success Rollouts
Trajectories where a baseline policy successfully completed the task.
Baseline Failure Rollouts
Trajectories where the baseline policy failed — useful negative signal for post-training.
Human-in-the-Loop Data
Human interventions, corrections, and preference labels collected during baseline rollouts.
All four components are intended for offline training in Phase 1. Hosted on Hugging Face Datasets.
Step-by-step instructions for joining the challenge — including environment setup, dataset access, baseline reproduction, evaluation protocol, and submission format — are hosted in our GitHub repositories. Reference code and starter scripts are provided so teams can get up and running quickly.
Based on Phase 1 results, the top 3 teams advance to Phase 2. During this round, selected teams collect rollouts by deploying their own policies and use those rollouts to further improve performance via post-training.
Top 3 teams selected from Phase 1 ranking
Iterative rollout collection supported by the organizers
Final policies presented at the workshop @ RSS 2026
Workshop @ RSS — team presentations and panel discussion
Awards
💰$20,000+Total Prize Pool
🥇
1st Place
$5,000
+ Bimanual YAM Robot
🥈
2nd Place
$3,000
🥉
3rd Place
$2,000
🏅
4th – 10th
$500 each
Top-performing teams will also be invited to give a presentation at the workshop.
Practical issues, lessons, and experiences from the challenge will be compiled into a technical report shared with the community.
Leaderboard
Phase 1 evaluation results across both tracks. The pi05 baseline is provided as a reference; team submissions will populate the table during the evaluation window.
#
Team / Model
insert-mouse-battery
tower-of-hanoi-game
seal-water-bottle-cap
Average
Score
SR
Score
SR
Score
SR
Score
SR
—
Baseline
pi05
87
80%
47
45%
47
35%
60.3
53.3%
#
Team / Model
insert-mouse-battery
tower-of-hanoi-game
seal-water-bottle-cap
Average
Score
SR
Score
SR
Score
SR
Score
SR
—
Baseline
pi05
—
—
—
—
—
—
—
—
Score: progress score per task (higher is better). SR: success rate over rollouts. The final ranking uses the Average column.
Registration
Team registration is now open. Please fill in the Google Form below with your team information and track preference. Submission instructions will be shared with registered teams.
