WAM-bam-thank-you NVIDIA: Long-WAM gives robots vastly better skills
Apparently, context matters. Giving a robot policy 38 seconds of visual history instead of none lifts its success rate from 63.3% to 78.7% on complex manipulation tasks, according to new research. That's a 15-percentage-point jump, and it comes without changing the robot hardware, the task setup, or the training data.
The catch, and the insight, is that you can't get that gain by just feeding a longer video sequence into a standard model.
The improvement only materializes if the underlying world model was pretrained autoregressively. A bidirectionally pretrained model, given the same extended context, shows essentially no benefit from the extra history.
That's the core finding of Long-WAM, a paper from NVIDIA, MIT, and the University of Hong Kong published on arXiv two days ago. The authors include Jim Fan and Song Han, two of the more prominent names in robotics AI research.
Long-WAM stands for Long-context World-Action Models. It's a robot control architecture built to actually use long visual context.
Why context length matters for robots
A robot trying to complete a multi-step task, say, assembling a part, sorting objects by type, or stacking cups that are moving, needs to keep track of what it has already done. Without memory, it has to infer task state from the current frame alone.
That works for simple, short tasks, but it falls apart for anything that unfolds over time or requires understanding how the world changed in the last few seconds.
The obvious solution is to give the model more frames to work with. The problem is that most robot foundation models are pretrained bidirectionally (non-autoregressively), similar to how BERT processes text: attending in both directions simultaneously. Those models show virtually no improvement when you extend the context window.
Long-WAM's backbone, called LongLive2.0-Robot, is pretrained autoregressively on roughly 10,000 hours of robot and egocentric video.
Autoregressive pretraining means the model learns to predict each frame given only the frames that came before it, the same direction causality actually works in. When you extend the context window at inference time, the model knows how to use that additional history.
The architecture that makes real-time possible
The engineering challenge is latency.
Processing 38 seconds of video at inference time sounds slow. On a real robot, you need actions fast enough to keep up with the physical world.
Long-WAM solves this with asymmetric attention: video tokens never attend to action tokens, which means the video stream can be processed asynchronously, in parallel with action generation rather than before it.
Combined with a streaming VAE encoder with CUDA-level optimization, the system hits 107.4 milliseconds per action chunk on an RTX 5090, and meets real-time requirements on a Jetson AGX Thor, the kind of edge hardware that could realistically ship inside a robot.
The benchmark results
The paper reports results across four benchmarks:
- LIBERO-Long: 99.5% average success
- RoboTwin 2.0: 94.4% average
- RoboCasa GR-1 with extended context: 78.7%, up from 63.3% with no history
- RoboCasa365 with GPT-6 Astra on 50 diverse tasks: 54.4%
(Note: these are company and author-reported figures, not independently reproduced.)
Long-WAM is compared directly against pi0.5 (Physical Intelligence's current leading manipulation model) and Fast-WAM, a prior model in the same research lineage. Long-WAM reports best results on GR-1, LIBERO-Long, and RoboTwin 2.0. The pi0.5 comparison is particularly notable: physical AI foundation models have become the benchmark to beat, and a new NVIDIA result that outperforms pi0.5 will get attention from the teams building on top of it.
Real robots, not just simulation
The paper includes real-robot results on a Unitree G1.
The task: dynamic cup stacking, where the target objects are moving rather than static. Success rate across 19 of 20 trials: 95%. The long context is doing real work here. A model without history would struggle to track a moving cup across frames. A model with 38 seconds of visual memory can.
The Unitree G1 result matters for context.
Humanoid robot deployments are accelerating, and the bottleneck is increasingly not the hardware, it's the control policies. A result that shows a 15-point success rate improvement from architecture changes alone, without new hardware or more training data, is the kind of thing that compounds across every platform that adopts the approach.
What this is not
Long-WAM is an arXiv preprint, not a peer-reviewed publication. The results haven't been reproduced by independent labs.
The real-world Unitree G1 tests cover a narrow task set. A 95% success rate on dynamic cup stacking is impressive; but we need more. The LIBERO-Long and RoboTwin benchmarks are well-established, but 99.5% on a simulation benchmark has a history of not surviving contact with physical environments.
The paper also doesn't address what happens on hardware with less compute than an RTX 5090 ... a real challenge for cheaper robots.
The Jetson AGX Thor results are promising, but real-world deployment headroom at 107ms per action chunk depends heavily on the task dynamics.
The pretraining insight is the real finding
It would be easy to read Long-WAM as a paper about context length, but it's actually more about pretraining. The finding that autoregressive pretraining is a prerequisite for context scaling to work is the kind of result that reshapes how the next generation of robot foundation models gets built.
Every lab currently training a generalist robot policy on video data has to decide: autoregressive or bidirectional? Long-WAM gives a concrete answer, and that's useful even for teams not adopting the full Long-WAM architecture.
Code and model weights are available at github.com/nvlabs/LongLive.
Given that this is coming out of NVlabs with open weights, the answer to "who will test whether this holds up across platforms" is: a lot of people, and soon.
Sources
Written by
Head of Marketing at NEXT Events & Conferences, responsible for Power Transmission & Distribution Technology Expo and Desalination Technology Expo. And, now, for Humanoid Technology Expo.