Thanks for the feedback! To clarify, it’s 5 trials/task across 10 tasks.
We’ll update to 15 trials/task. With 7 policies already evaluated, that’s ~6–8 hours per policy.
That's challenge of real-world evaluation, but we’re committed! @JiahuiZhang__32@alexhuang1029@maylynn_e
5 trials on this leaderboard is a highlight reel. I'd want intervals, cycle time, jam recovery, overrides and intervention data.
irvlutd.github.io/VLAReplica/#le…
@bidhan That's exactly why RobotReplica exists. One standardized evaluation so results aren't tied to whatever robot happens to be under each model. Check it out! :)
robot-replica.org
We finished evaluating MolmoAct2 on VLA-Replica! It shows better OOD performance than π0.5. Check out the details from @maylynn_e@alexhuang1029@JiahuiZhang__32 below 👇
We’ll keep benchmarking frontier models for robot manipulation in the real world. Next up: GPT-6 Astra!
Finished fine-tuning MolmoAct2 on our VLA-Replica SO-101 setup, using 50 demonstrations per task!
One result I find interesting: it outperforms π₀.₅ on OOD by 10%.
Check out our leaderboard:
irvlutd.github.io/VLAReplica/#le…pic.x.com/27qkS6Swfh
Really cool to see evaluation done in such a systematic and reproducible way. Standardized evaluation is one of the hardest problems out there right now!
We finished evaluating MolmoAct2 on VLA-Replica! It shows better OOD performance than π0.5. Check out the details from @maylynn_e@alexhuang1029@JiahuiZhang__32 below 👇
We’ll keep benchmarking frontier models for robot manipulation in the real world. Next up: GPT-6 Astra!
ID: MolmoAct2 reaches 46% success, compared to 54% for π₀.₅, ~30% for SmolVLA, and ~30% for π₀
irvlutd.github.io/VLAReplica/sce…
OOD: MolmoAct2 reaches 45% success across 8 VLA-Replica-OOD tasks, compared with 35% for π₀.₅, 30% for SmolVLA, and 30% for π₀
irvlutd.github.io/VLAReplica/sce…
Finished fine-tuning MolmoAct2 on our VLA-Replica SO-101 setup, using 50 demonstrations per task!
One result I find interesting: it outperforms π₀.₅ on OOD by 10%.
Check out our leaderboard:
irvlutd.github.io/VLAReplica/#le…pic.x.com/27qkS6Swfh
We finally got a chance to test MolmoAct 2 on our VLA-Replica setup! @DJiafei@hq_fang
Here’s one successful “zero-shot” case from our side :) @maylynn_e@alexhuang1029@JiahuiZhang__32
We put “zero-shot” in quotes since we don’t know what data this MolmoAct 2 model was trained on.
Two preliminary observations from our tests:
• It seems quite sensitive to object color. Detailed instructions with color can help.
• It struggles with precise grasping of small objects, such as a block.
525 Followers 4K FollowingLLM Post-Training/ Recursive Self-Improvement for Reasoning
Co-founder/ MTS @ReasonCoreAI
PhD @UCLA, Bachelors @iitbombay
Led a team of ~500 PhD researchers
8K Followers 1K FollowingAssistant Professor at @NUScomputing | Robotics & AI PhD @uwcse | Host of @RoboPapers | Ex-@allen_ai robotics lead, @NVIDIA | Magician |
my opinion is my alone.
7K Followers 1K FollowingI do research in robotics and computer vision @IRVLUTD, Ex- research scientist @NVIDIA, postdoc @uwcse and @StanfordAILab, phd @UMich, all views my own.
8K Followers 1K FollowingAssistant Professor at @NUScomputing | Robotics & AI PhD @uwcse | Host of @RoboPapers | Ex-@allen_ai robotics lead, @NVIDIA | Magician |
my opinion is my alone.
7K Followers 1K FollowingI do research in robotics and computer vision @IRVLUTD, Ex- research scientist @NVIDIA, postdoc @uwcse and @StanfordAILab, phd @UMich, all views my own.