Marius Memmel (@memmelma) profile picture

Robotics PhD student @UW and intern @amazon FAR, previously intern @NVIDIA SRL, intern @Bosch_AI, @EPFL, @TUDarmstadt, @DHBW

Seattle, WA
Joined April 2021
Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)Photo by Marius Memmel (@memmelma)
Want to turn any VLM into a better reward model for robotics? (yes, even Astra) Introducing D&A: decoupling description and assessment improves any VLM family for zero-shot robot rewards. D&A scales with improvements across GPT, Gemini, Claude, Qwen, ... It enables zero-shot RL and filtered IL, improves code-as-policies via rich language feedback, and achieves SOTA on three benchmarks. Website: weirdlabuw.github.io/dna
14
48
15
258
29,437
At #CoRL2026, many of you asked for best practices on using D&A as a robotic reward model 🧬 Here is a cheat sheet of tips & tricks to get cleaner rewards, faster inference, and a lighter hit to your wallet: 1. Inject as much prior knowledge as possible. D&A is designed to be general, but because it’s prompt-based, you can guide any stage of the pipeline: - Hardcode or cache object/entity names to ensure the VLM recognizes the objects in the scene and that they stay consistent across predictions. - Make language instructions oddly specific (e.g., "pick up the hook piece, insert it, then pick up the other one and hang it"). Better instructions = better predicted success criteria. - If you already have a rubric in mind, explicitly set the success criteria for more stable, less noisy rewards. - Use potential-based reward shaping if the reward magnitude feels noisy or inconsistent. - Average multiple runs per prediction if budget isn't a bottleneck. 2. Match frame density to your downstream task - Need fine-grained rewards (e.g., offline RL)? Annotate more frames per trajectory. - Just doing coarse subtask filtering? Annotate fewer frames. - Rule of thumb: Larger models handle 16–64 frames well. Smaller models hit their sweet spot at 8–16. 3. Debug through the language interface - Because D&A uses natural language between stages, inspect intermediate outputs to see where things break. → Missed an object? Reword the entity description. → Missed a subtle failure mode or subtask? Make the task instruction more specific. - Plot success vs. failure trajectories to verify you have clean reward separation. 4. Optimize inference speed & cost - Lower image resolution: Full resolution is rarely needed for strong performance. ⁃ Downsize the base model: Test smaller VLMs to find the best speed-to-accuracy trade-off for your task. weirdlabuw.github.io/dna
Want to turn any VLM into a better reward model for robotics? (yes, even Astra) Introducing D&A: decoupling description and assessment improves any VLM family for zero-shot robot rewards. D&A scales with improvements across GPT, Gemini, Claude, Qwen, ... It enables zero-shot RL and filtered IL, improves code-as-policies via rich language feedback, and achieves SOTA on three benchmarks. Website: weirdlabuw.github.io/dna
1
4
14
841
Marius Memmel retweeted
How about: Whole-body motion tracking for the Unitree G1 in under a minute :) XMimic trains one policy on all 40 LAFAN1 clips in 54 seconds on 4 H100s. 3.5x faster than BlinkMimic and 85x faster than SuperMimic. Code: github.com/nico-bohlinger/xm… Report: nico-bohlinger.github.io/pap…
16
44
8
333
22,226
Marius Memmel retweeted
Turns out that humans are far from optimal when intervening on robot behavior. This makes many algorithms for learning from human interventions either too conservative or incorrectly reliant on human optimality. @_pranavnt, @sriyash__ and team first analyze how people actually intervene on robots. Then they use this to a propose a simple RL algorithm - Q-OIL that uses these interventions to learn very quickly. Q-OIL proposes a simple idea - be optimistic on state-action pairs where people intervene. Think of it as - "even if the human isn't perfectly right, it's probably worth checking that out". This leads to fast improvement and a lot fewer human interventions for robots learning with RL in the real world. If you are doing RL in the real world using interventions, this should be a simple drop in for things like HiL-SERL etc. We've got lots of robot results + a fun interactive demo. Website: weirdlabuw.github.io/qoil/ Paper: weirdlabuw.github.io/qoil/pa…
Excited to share the first paper of my undergrad, Q-OIL, which we’ll be presenting at #CoRL2026 in Austin! Interventions provide a natural source of feedback for robots to learn from in the real world. But, humans intervene suboptimally, which needs to be accounted for in both simulation benchmarking and algorithms. We analyze human intervention data and measure intervention suboptimality in both timing and execution. Thus, rather than using interventions as supervision, we use them as soft exploration guidance. (1/8)
8
12
66
4,426
Marius Memmel retweeted
Excited to share the first paper of my undergrad, Q-OIL, which we’ll be presenting at #CoRL2026 in Austin! Interventions provide a natural source of feedback for robots to learn from in the real world. But, humans intervene suboptimally, which needs to be accounted for in both simulation benchmarking and algorithms. We analyze human intervention data and measure intervention suboptimality in both timing and execution. Thus, rather than using interventions as supervision, we use them as soft exploration guidance. (1/8)
15
37
9
160
19,486
Really interesting experiment by @SwannAiden! What if we asked a model to compute a robot arm's kinematics, but gave it the specs of a different robot than the one it knows (Franka)? The best models (GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon) ignored the specs and answered as a Franka Panda anyway. Their answers landed 15–21 cm from the Panda's and 80–96 cm from the correct ones. That looks like memorization of one robot, not general kinematics! Claude and I tried to make the insight a bit more digestible:
GPT-6 Astra's robotic capabilities are not emergent. It was trained on robotics specific data. So were GPT-6.1 Sol and Gemini Argon Here's how we know 🧵
3
4
1
19
1,736
Glad to finally see what Dieter has been up to at Ai2! Of course, D&A works on blocks too 😉 teapawt.net/memmelma/status/211762…
We’re introducing ToyBlocks, an open simulated tabletop for controlled experiments in robot learning. It enables robotics researchers to experimentally control training data & models, run targeted evals, & trace failures to their causes. 🧵 github.com/allenai/ToyBlocks
1
2
19
861
Marius Memmel retweeted
Can an off-the-shelf VLM be a reward model for robot learning, with no robot data and no fine-tuning? Asked directly, VLMs tend to assume the task succeeded and report progress on failed runs. Fine-tuned reward models such as Robometer do better, but are small open models that must be retrained for every new base model. D&A from the University of Washington and NVIDIA gets a better zero-shot reward out of any VLM with prompting alone, by keeping the task hidden from the model that narrates the video. How it works • First, every object in the scene is listed, relevant or not. A describer then narrates the whole video with these names, without knowing the task or which objects matter. • An assessor sees only the instruction and that text, never the video. It first writes an explicit success criterion, then scores progress from 0 to 1 at each sampled frame. • No training. With Gemini models a trajectory costs 1.5-6 cents and 24-58 s of API calls, so it suits RL and data filtering, not closed-loop control. Does it pay off? • On the authors' D&A-Bench (54 real trajectories, half of them failures), Gemini 3 Flash with D&A reaches Kendall τ 0.67, vs 0.57 for Robometer. The same prompts also lift Qwen and GPT models, the weaker ones most. • As the dense reward for online RL on a DROID Franka, it takes a π0 policy from 10% to 100% success on putting the fruit in the bowl in about 90 min. TOPReward rewards the wrong object instead. Caveats • With Gemini 3 Flash, Robometer still wins on its own benchmark (0.648 vs 0.576). The SOTA claims need GPT-6 Astra, where D&A adds only 0.02-0.04 over a plain prompt. • Each real-robot result is a single task, and the paper gives no trial counts. Code is out (MIT). By @memmelma, @imankitgoyal, @Jesse_Y_Zhang, @fox_dieter17849, @AnqiLi24, @abhishekunique7 and co-authors.
Want to turn any VLM into a better reward model for robotics? (yes, even Astra) Introducing D&A: decoupling description and assessment improves any VLM family for zero-shot robot rewards. D&A scales with improvements across GPT, Gemini, Claude, Qwen, ... It enables zero-shot RL and filtered IL, improves code-as-policies via rich language feedback, and achieves SOTA on three benchmarks. Website: weirdlabuw.github.io/dna
2
1
17
1,064
Wait, isn't this just BERT/Jev but with explicit CV- and robotics-specific outputs?
The era of System One Vision Model is here! Introducing VFlash: parsing the visual world in parallel at unmatched speeds for Physical AI. Most embodied prediction tasks like object localization, robot trajectory planning, and dense 3D motion are fixed-length float vectors. For example, a bounding box is always 6 tokens, a 10-waypoint robot arm trajectory is 54: we know these beforehand. Instead of decoding one token at a time, we can output the entire geometric shape in one go. VFlash enables up to 12x faster action decoding and a 5.1x overall robot inference speedup on edge GPUs like the Jetson Thor and DGX Spark. What’s surprising is that it ends up even more accurate than the VLM backbone, with up to 61% lower 3D trajectory error. The intuition is that VFlash avoids the cascading errors of autoregressive decoding, where every coordinate is written in ink and every later point loses the chance to backspace. Everything that moves will be autonomous, and everything autonomous will need to think on its own chip. The next inference revolution is physical reasoning at the speed of reflex. And as always, it’s open paper, open source! Link in thread:
3
1
16
3,101
Marius Memmel retweeted
Really excited to share Catch It All!, my first paper as last author! In this work led by @Kehlani_Fay @arth_shukla we develop a system for catching diverse objects using a single egocentric camera 🧵👇
4
24
2
102
5,562
Really interesting work! Reminds me of SimEX, but with storing sim experience in the model's weights instead of reusable skills: teapawt.net/JiahengHu1/status/2107…
Astra made robot control look like a job for frontier LLMs. But small VLMs already know what to do. They just don't know how. Turns out, simulation is all you need. Meet ZAP. No real robot data, no frontier teacher. Matches Astra on real tasks, 3× faster, 17× cheaper. 🧵 (1/11)
8
919
💪 We teach humanoids where to move and how hard to apply force, using human demonstrations collected without a robot. Introducing 🌺 FLoRe: Learning Force-Reactive Humanoid Loco-Manipulation. Our policy learns to predict actions and forces for each hand. A hybrid compliance whole-body controller tracks these commands while the robot walks and balances. 🧵 (1/6)
19
44
8
228
13,234
Chicken costume, Lederhosen, Labubu... how far can we push D&A?? Further than we thought! 🐔🇩🇪 Fun fact: the model outright refused the descriptions: ❌ "the chicken" (worked: "chicken costume") ❌ "the German" (worked: "person in Lederhosen") Well done @OpenAI alignment team 😉
Want to turn any VLM into a better reward model for robotics? (yes, even Astra) Introducing D&A: decoupling description and assessment improves any VLM family for zero-shot robot rewards. D&A scales with improvements across GPT, Gemini, Claude, Qwen, ... It enables zero-shot RL and filtered IL, improves code-as-policies via rich language feedback, and achieves SOTA on three benchmarks. Website: weirdlabuw.github.io/dna
10
446
So close! Loved seeing the successful video on the website @zhiyuan_zhou_ teapawt.net/memmelma/status/211762…
We always talk about models in robot learning, yet data quality is just as important (if not more!). How do we collect good data? It turns out RL can help you *during* data collection! If you're collecting demos, ✨Data Inoculation✨ is a simple addition to collect better data!
5
363
We always talk about models in robot learning, yet data quality is just as important (if not more!). How do we collect good data? It turns out RL can help you *during* data collection! If you're collecting demos, ✨Data Inoculation✨ is a simple addition to collect better data!
17
16
7
190
34,168
What if your robot is a lamp? D&A still gotchu haha teapawt.net/memmelma/status/211762…
Excited to share that we've raised $11M in pre-seed funding to bring Lume to more homes. Our lamp that does your chores, was also recently named one of TIME's Best Inventions of 2026! Grateful to everyone who believed in @bySyncere early. Back to shipping.
5
498
Marius Memmel retweeted
I wasn't there for the big launch, but I saw the grinding from @Daphne__Chen @sidharthtalia @arc_hitter last week at 11pm in the lobby on Thursday night. The amount of work it takes to get a multi-hour social demo out the door is monumental! Major respect
Our new robot OpenDubs worked a 6 hour shift autonomously handing out candy to trick-or-treaters for Halloween this year @uwcse 🤖🍬🎃 Thanks to the many folks who made this fun side project into a reality 😄 We took quite the full stack approach to this endeavor - literally building this robot from the ground up in just under a month and a half! @sidharthtalia @arc_hitter @siddhss5 @uw_robotics Stay tuned to see more of what OpenDubs can do! 🦾
3
13
772
Marius Memmel retweeted
Back in ~Dec 2025, I really wanted to work on mobile manipulation, but robots were either too expensive or not performant enough. So we built our own! Excited to finally share something I’ve had so much fun working on: 1. Pixel: a co-designed full stack system for learning mobile manipulation. 2. Pixel-UMI: environment aware whole body manipulation without robot data. w/ @JasonJZLiu (1/N)
26
55
24
347
41,896
Marius Memmel retweeted
VLMs grading robot videos: "great progress!" The robot: holding an ice cream cone instead of the banana. D&A's fix is blind grading: describe the video without seeing the task, then grade it. 10% → 100% on real-world RL. Works with GPT, Gemini, Claude and Qwen 👏 @memmelma
Want to turn any VLM into a better reward model for robotics? (yes, even Astra) Introducing D&A: decoupling description and assessment improves any VLM family for zero-shot robot rewards. D&A scales with improvements across GPT, Gemini, Claude, Qwen, ... It enables zero-shot RL and filtered IL, improves code-as-policies via rich language feedback, and achieves SOTA on three benchmarks. Website: weirdlabuw.github.io/dna
1
1
5
236
This is awesome!! Take a (partial or full) video of any robot performing any task along with a text description, pass it through D&A, and you get an off-the-shelf reward model, with no training! Here it is on the SGS locomotion task, and it's super cool how D&A only gives extra reward when the goal is visible in the image Great work from @memmelma and the team!!
Want to turn any VLM into a better reward model for robotics? (yes, even Astra) Introducing D&A: decoupling description and assessment improves any VLM family for zero-shot robot rewards. D&A scales with improvements across GPT, Gemini, Claude, Qwen, ... It enables zero-shot RL and filtered IL, improves code-as-policies via rich language feedback, and achieves SOTA on three benchmarks. Website: weirdlabuw.github.io/dna
5
3
25
1,754
Marius Memmel retweeted
Super exciting to see D&A in action on our RFS robot manipulation videos! 🤖🔥 Really cool to see the robot behaviors analyzed and visualized in this way. Awesome work! 🚀
Want to turn any VLM into a better reward model for robotics? (yes, even Astra) Introducing D&A: decoupling description and assessment improves any VLM family for zero-shot robot rewards. D&A scales with improvements across GPT, Gemini, Claude, Qwen, ... It enables zero-shot RL and filtered IL, improves code-as-policies via rich language feedback, and achieves SOTA on three benchmarks. Website: weirdlabuw.github.io/dna
3
12
742