createnotfind

robots and neural nets.

[createnotfind @createnotfind]

@zipengfu @sippeyxp @chenwang_j 请问后面会放出更详细的技术报告或者 blog 吗

[createnotfind @createnotfind]

Three steps in scaling robot learning:

EgoScale: scale actions extracted from human videos.
Dyna-2: scale egocentric video pre-training.
Rhoda AI: scale general web-video pre-training.

Less robot-specific supervision, yet scaling still improves downstream robot performance.

[Tongzhou Mu 🤖🦾🦿 @tongzhou_mu]

Does scaling pre-training on general web video improve a complex manipulation task in real deployment?
We scale model size and pre-training compute, and test on one industrial task.
Yes. The better a pre-trained model predicts web video, the better its post-trained policy. 🧵

[createnotfind @createnotfind]

@roeiherzig This approach appears to rely primarily on LLMs’ numerical pattern-matching capabilities rather than spatial understanding.

[createnotfind @createnotfind]

@dimentary I'm curious that do you use the codex directly to complete this task or you just let the gpt6 model input images and joint angles and output joint angles similar to Inspect Robots

[createnotfind @createnotfind]

@k7agar what if give it an example (video demonstration)?🧐

[createnotfind @createnotfind]

Do you think GPT-6 has already been trained on robot control tasks?

[Jay Chooi @chooi_jeq]

GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower cost. 🧵

[createnotfind @createnotfind]

@BruceLeehimself @gkxspace 哈哈哈，实际上人可以是吧

[createnotfind @createnotfind]

@BruceLeehimself @gkxspace 主体在境外也算国内违法吧

[Robots Digest 🤖 @robotsdigest]

N⁰-Foundation: A foundation for tactile-enabled embodied manipulation.

Vision alone isn't enough for many real-world manipulation tasks.

Deformable objects, precise assembly, delicate force control, and sustained contact all require information about what is happening at the point of interaction.

N⁰-Foundation builds the infrastructure, data, representations, and benchmarks needed to make tactile learning scalable.

[createnotfind @createnotfind]

@junfanzhu98 @saturdayrobotic @KDharmarajan123 @UCBerkeley @Stanford @aurorafeng_01 @jerryhuang01 @RoboticsCtr very insightful sharing, would be great if there were videos！

[createnotfind @createnotfind]

@chris_j_paxton looks like motion planning

[createnotfind @createnotfind]

@lucasmaes_ @seohong_park very long context window

[createnotfind @createnotfind]

@Pushkar11420566 Isn’t GEN’s pretraining data already collected with UMI-like handheld grippers, rather than ordinary egocentric video?

[createnotfind @createnotfind]

How do you train on 1M+ hours of video?

Dyna-2:

• MCAP + H.264 large GOP → 68% less storage
• topic-group chunking → 2.9× faster reads
• Airflow+ bin packing → 31× faster ingestion
• SQL manifests → startup 48h → <1min
• Alluxio + local NVMe → ~98% GPU util

[Dyna Robotics @DynaRobotics]

When people talk about robotics, they usually talk about models, data, or hardware. Few people talk about the infrastructure that lets you iterate on all three quickly. Today we're publishing how we trained Dyna-2 on over 1,000,000 hours of egocentric video, repeatably. At this scale, most of what worked at ten thousand hours did not hold up:

• ingestion throughput was capped at 14,000 episode-hours per week — a million hours would have taken over a year
• building a training manifest took 48 hours before a run could even start
• reading a petabyte from cloud storage during training left GPUs exposed to latency and packet loss

🧵

[createnotfind @createnotfind]

How do you train on 1M+ hours of video?

Dyna-2:

• MCAP + H.264 large GOP → 68% less storage
• topic-group chunking → 2.9× faster reads
• Airflow+ bin packing → 31× faster ingestion
• SQL manifests → startup 48h → <1min
• Alluxio + local NVMe → ~98% GPU utilization

[createnotfind @createnotfind]

@KeivalyaP great advice！ need more～

[createnotfind @createnotfind]

@yacineMTB curious about the bug

[createnotfind @createnotfind]

@ChenTessler could you like to share your exp setting? i tried last night and it seems like this. I sweep lr and select the best for each optimizer

[createnotfind @createnotfind]

fantastic

[Chen Tessler @ChenTessler]

I'm late to the Muon train, but damn this is fast.
Seeing 2-3x speedup on DeepMimic-style training (RL).

Need to do a speed run competition on the bones-seed dataset.
Research will move so fast once we can solve for huge datasets in 1 hour.

[createnotfind @createnotfind]

@vargastartup Does the "physics" here refer to skeletal 3D keypoints? I've heard that acquiring their ground truth is extremely difficult, which makes large-scale deployment challenging. Additionally, retargeting introduces errors and thus degrades model performance.