Pinned Tweet
There’s a discussion going on rn about two recent robotic reward models: TOPReward⛰️ and Robometer🌡️
Which one is better? It depends entirely on your objective!
Here is a deep dive into the conceptual differences, strengths, and weaknesses of both. 🧵👇
Some years ago, @fox_dieter17849 suggested exploring dexterous sphere handling. I fed his video to @OpenAI's GPT-6 Astra and asked it to convert human motion into a robot trajectory featuring the @SharpaRobotics hand.
Here's what it did 🧵
Main takeaway: GPT-6 Astra loves CaP and yields open-loop trajectory reconstructions rather than reactive, generalizable policies across varied spheres or initializations.
However, this setup might be a good starting point to warm-start off-policy RL replay buffers or sampling resets for RL, similar to:
Learning Montezuma's Revenge from a single demonstration [1] → OmniReset [2] → SGS [3] →AlphaNIST [4]
[1] arxiv.org/abs/1812.03381
[2] arxiv.org/abs/2603.15789
[3] arxiv.org/abs/2610.12465
[4] arxiv.org/abs/2610.13354
# prompt 1
"
Input:
video of a human handling a sphere [video].
Task:
reconstruct the scene and generate the sequence as a robotic trajectory in simulation using the Sharpa Wave hand (attached to a robot arm) [github.com/google-deepmind/m…].
Output:
script (reproducible) generating a full dynamic sequence resembling the exact sphere and hand movements from the video with correct contact dynamics. you're evaluated by a human roboticist inspecting the two videos.
Rules:
system: nvidia T4 GPU
only read and write to your home directory
only read images <3mb in context to avoid crashing the session
"
codex (GPT 6 Astra high) would crash sometimes when reading large images because they blew up its context
after 3 days, it was still trying to refine the exact rotations of the ball to match the logo on it exactly, so I decided to call it
# prompt 2
"
this is taking too long. finish up with what you have. then re-render the full rollout to look realistic.
"
Marius Memmel ✈️ CoRL 2026 retweeted
Can and should a general-purpose model serve as a robot policy?
In Astranomy, we study GPT-6 Astra for general pick-and-place manipulation. Astra sets a new SOTA on RoboLab-120, but real-time robot control remains an open challenge.
Full tech report:
research.nvidia.com/labs/srl…
Marius Memmel ✈️ CoRL 2026 retweeted
What to remember in the past to act—spatial, temporal or motion details?
We often assume what’s useful, but that’s hard to know in advance.
Meet IMPRINT 🧠: retain rich memory first, then learn to use it for control. Compact latents keep inference fast as history grows. 🧵 1/10
Marius Memmel ✈️ CoRL 2026 retweeted
(1/6) From @GoogleDeepmind’s AlphaGo to AlphaZero, we saw machines find moves that reached beyond what we’d imagined for them. No one taught them those strategies. They emerged from reinforcement learning: playing, failing, and trying again under a single binary task reward for success or failure. Inspired by them, we couldn’t help but wonder: what might a robot discover if we gave it that same freedom to learn?
That question led us to AlphaNIST. Starting tabula rasa, with that same binary reward plus only a simple control cost, a single policy solves all 20 @NIST assembly tasks at a 98.2% success rate in simulation. More surprising still, AlphaNIST can complete 10 different tasks in a row, receiving that reward only once, after all 10 are done.
Just as AlphaGo’s strategies surprised its creators, AlphaNIST discovered dexterous strategies we couldn’t have come up with ourselves.
Paper: arxiv.org/abs/2610.13354
Website: alpha-nist.github.io
GPT-6 Astra unlocks a whole new level of simulation scaling! It generates diverse environments, designs and solves tasks, and deploys seamlessly across embodiments.
Below: 25 unique environments, 25 multi-step tasks, and 5 robot embodiments generated in under 24 hours with just 3 generic prompts!
🧵
Huge thanks to @kevin_zakka and collaborators for developing MuJoCo Menagerie [3], and to @seohong_park for key insights on generating human-like demonstration data [4]!
References:
[1] arxiv.org/abs/2602.11337
[2] arxiv.org/abs/2603.04356
[3] github.com/google-deepmind/m…
[4] seohong.me/blog/behavioral-c…
Marius Memmel ✈️ CoRL 2026 retweeted
🚀 Looking for a reliable Franka controller with seamless sim-to-real transfer?
Meet FrankaTwin: github.com/tsrobcvai/frankat…
🎮 Joint/Cartesian position & impedance + gripper
🎯 System ID: 3.6 mm EE / 28 mrad joint RMSE (sim vs. real)
Remember URDFormer? The vision was simple: image in, URDF out.
Fast-forward 2.5 years, and GPT-6 Astra can solve this task zero-shot.
This is a massive win for the community: a scalable and cheap way to generate high-quality simulation assets!
The crazy part? This worked on the very first try with the dumbest prompt I could come up with:
"You are a robotics engineer. Convert the scene shown in the image a single valid URDF description."
GPT-6 Astra (Max)
While the original URDFormer architecture is now obsolete, the premise has aged like fine wine 🍷
urdformer.github.io/
@ZoeyC17 @aaronwalsman @abhishekunique7
Marius Memmel ✈️ CoRL 2026 retweeted
We are giving out $500k and free YAM arms to academic teams to design dexterous manipulation benchmarks and run real-world evaluations.
General-purpose robots could soon be in homes and data centers. Society should know what robots can do before they get there.
ALT Dark teal gradient with white text. Small label "Expression of Interest" above the headline "Call for Open-Source Robotics Benchmarks" and the line "We are funding academic teams to design open-source robotics benchmarks for dexterous manipulation and to run real-world evaluations."