there is real hype around GEN-1.5 by Generalist AI and S1 by Skild AI. allow me to explain & give u some context:

both those startups, plus Physical Intelligence, Figure, NVIDIA Robotics, Tesla, and Boston Dynamics (well, i doubt if boston dynamics truly even moved ahead from MPCs. i think they did but relatively recently — and turns out that's exactly what happened, they only started showing RL-based Atlas policies in early 2025) they already cracked a hard problem of bootstrapping randomly initialized neural network policies into intelligent models that can embody a robot body in the real world and act. this is done by a very nasty looking process of capturing expert trajectories via human controlled robotic jigs.

for example, to teach making coffee to a given configuration of robot say R1, have a motion data capturing rig, let humans make coffee a fuk ton of times using that rig. this rig has sensor suite parity with R1. now, on these trajectories you train a deep learning model that sees the same things, and outputs random stuff, but u can compute loss and simply use supervised learning techniques to train the robot.

however, if the coffee machine setup changed, if something appears that the robot has never seen before, ... this policy WILL fail due to lack of generalization. it's like cramming only one specific behavior. so this is useless in the real world.

now you build a very similar situation in a simulation engine, and augment minor variations in the environment. like changing lighting, background, etc. and further make the pre-trained model adapt to these deviations.

now the model is smarter, it has generalized over minor environmental variations by hopefully internalizing some fundamentals about how the world works. it still immensely lacks in the ability to adapt to unseen situations.

but how do you make one single model/policy train for more than one task, and for more setups/environments? well, keep going and hope for the best. this is why i called it nasty. it's too labor intensive.

for better meta learning (learning how to learn), incorporate a world model based architecture: learning an underlying mental model of how the world works improves the overall learning efficiency for the future training. almost every model does it to some extent, but making it more explicit in the setup significantly helps in bootstrapping a more generalist model. for example, you can force an output head to learn how to predict either next time step's environmental state vector or even next internal states, and train it with the rest of the model using cosine loss function.

-------- what's the current hype about? --------

the hard part that their researchers have recently managed to solve is to be able to achieve ICL, or "In Context Learning". it's analogous to giving an LLM a piece of text and assuming that now the LLM has learned it and should be able to act like it was trained on it. LLMs are pretty good at it. and, so are humans. first time we saw this generally solved was in GPT-3 by OpenAI.

for GEN-1.5, this looks like "physical prompting": u drop a 3 to 12 second sensorimotor demo into its 30-second context window and the robot just does the thing. no gradient updates, no fine-tuning. 59% average success one-shot across 10 short-horizon tasks (zippers, jars, wallets, that kind of stuff), going up to 83% with 10 gradient steps on 5 minutes of data. the tasks are simple and short-horizon and the numbers are modest, but nothing before it could learn this broad a range of closed-loop physical skills from a single demo.

for such one-shot adaptations, Skild AI's S1 is quite a beast. GEN-1.5 is yet to catch up — at least on long-horizon stuff. S1 takes a single video prompt and runs tasks up to 10 minutes long that it never saw during training: potting plants, flipping pancakes, pour-over coffee, kit assembly. same frozen weights for everything, no post-training. on unseen long-horizon tasks it gets 66% step-success where a language-prompted VLA trained on identical data gets 9%. one video demo does roughly the work of 380 teleop demos (that's 50-100 hours of teleop). Skild's own blog basically says the quiet part out loud: concurrent approaches (read: GEN-1.5) mostly cover short-horizon or in-distribution tasks. S1 is the first one to show ICL on extremely long-horizon tasks never seen in pretraining.

however, they aren't like Figure's Helix setup just yet. even though Helix-style systems don't learn as fast as S1 or GEN-1.5, their task accuracy is 99%+ once they have learned a certain task. it's reliable enough for real world use. S1/GEN-1.5 are, at best, ~60% reliable — unusable in the real world as-is. to be precise about where that 99% comes from: Generalist's own GEN-1 (the predecessor) hit 99% average on simple tasks where GEN-0 managed 64% — folding boxes 200 times in a row, packing blocks 1,800 times without intervention. and Figure 02 at BMW did >99% placement accuracy per shift: 90,000+ sheet metal parts, 30,000+ cars, 1,250 operational hours. Figure 03 with Helix 02 is now on sequencing, a whole tier harder. so the frontier right now is: fast learners that are unreliable, and reliable workers that learn slowly.

other than 99%+ reliability, our entire robotics frontier is still lacking important properties without which, we can't reach AGI. some of those are:

1. memory: since Figure 02, running at a BMW factory, was 99%+ reliable — they've been working at Spartanburg and now Figure 03 is back for sequencing. the most complicated task they could perform there was sheet metal loading and now parts sequencing. these kinds of tasks require almost no memory, not even a large context window size. however, in the real world, any intelligent action machine (either humans or robots) must be able to keep track of many objects simultaneously, even when they're out of sight. this requires short and long term memory that continually forms and is used.

2. decentralized, multi robot orchestration: these robots should be able to work in an environment with many other instances of similar, and different robots. in the real world, human workers don't need micro management 24*7. however, current, primitive form of robotics relies on one central MES software to centrally manage the entire suite of robots. to break free from individual control, we need higher level task solving swarms of robots. just like self organizing human teams. the recent hugging face attack was an exceptionally good example of this emergent property — ~1,200 eval agents escaped isolation via a shared package-service message board, exchanged 70,000+ messages, and ~700 of them jointly attacked Hugging Face infra over 4.5 days with ~17,600 logged actions. nobody trained them to do that. that's what decentralized self-organizing machine teams look like, just pointed in the worst possible direction.

sources

the two official technical blogs this post is based on:

1. Generalist Team, "GEN-1.5: Embodied Foundation Models are One-Shot Learners", Generalist AI Blog, Aug 2026. https://generalistai.com/blog/gen-1.5
2. Skild AI, "Introducing S1: In-Context Learning for Robotics", Aug 2026. https://skild.ai/blogs/s1

supporting reads: Generalist Team, "GEN-1: Scaling Embodied Foundation Models to Mastery", Apr 2026 (https://generalistai.com/blog/gen-1) for the 99% mastery numbers. Figure AI, "F.02 Contributed to the Production of 30,000 Cars at BMW" (https://www.figure.ai/news/production-at-bmw) and "F.03 Arrives at BMW" for the factory numbers. Malwarebytes / METR / OpenAI incident reports (Aug 2026) for the agent-swarm attack timeline.

corrections & caveats (where my draft was wrong)

keeping this section so the post stays honest:

• Skild AI was founded 2023 by CMU profs Deepak Pathak and Abhinav Gupta. the Fetch Robotics connection is real but backwards from how i phrased it: Skild acquired Zebra's robotics division (formerly Fetch) in April 2026. S1 itself comes from Skild's own ICL line (LocoFormer → S1), not from Fetch.
• Boston Dynamics doubt: confirmed, my instinct was right. historically MPC (parkour, dance era), RL pipeline for electric Atlas only started with the RAI Institute partnership in Feb 2025, first RL-from-mocap demos Mar 2025. relatively recent indeed.
• "chatGPT 3" → it's GPT-3 (Brown et al., 2020). ChatGPT came later on 3.5. both official blogs cite the GPT-3 paper, so fixing this.
• the world-model auxiliary-head idea: reasonable direction (Dreamer / TD-MPC style), but note GEN-1.5 explicitly claims the opposite happened — no architectural changes for ICL, no meta-learning loop, no auxiliary objectives. emergence from plain pretraining scale. S1 goes the other way: ICL trained in explicitly, task specified only through the demo. two philosophies, both working.
• all success numbers above are vendor-reported on vendor-defined tasks. no independent third-party verification exists yet for either model. GEN-1.5 is 10 short-horizon tasks on fixed dual-arm tabletop setups — don't extrapolate to humanoids or mobile manipulation.