The Catastrophic Forgetting Problem Nobody Wants to Talk About

There's a related paper on humanoid robots that's more honest about the underlying problem. HumanoidMimicGen, described in another arXiv preprint, is trying to automatically generate training data for humanoid loco-manipulation (walking while manipulating objects, which turns out to be really hard). The researchers found that policies trained with their generated data outperformed those trained only on real-world data by 20%. That's a meaningful number! But read the fine print: this is in simulation, on their own benchmark, and the humanoid action space is so high-dimensional that nobody really knows how well any of this transfers to physical hardware.

The Touch Problem

One of the more interesting papers this week tackles something most robot AI research conveniently ignores: touch. A team working on dexterous manipulation introduces what they call Center-of-Pressure (CoP), a tactile representation that tries to preserve dense contact information while remaining robust enough for sim-to-real transfer. According to arXiv, policies using this representation achieved zero-shot sim-to-real transfer on a multi-fingered hand for tasks like peg-in-hole insertion and ball balancing.

This is actually impressive if it holds up. The reason most robot learning research ignores touch is that tactile sensors are wildly inconsistent between simulation and reality. The paper claims they solved this with a sensor calibration scheme based on differentiable dynamics, which lets them estimate sensor orientations without ground-truth force measurements. Whether this works outside their specific setup remains unclear, but at least they're attacking a real problem instead of pretending it doesn't exist.

The key findings from this week's research, for those keeping score:

• VLA models suffer catastrophic forgetting when learning sequential real-world tasks
• Experience replay helps but isn't sufficient, specific implementation factors matter more than the technique itself
• New simulation frameworks can generate training data 25 frames in 2.3 seconds, but real-world transfer remains the bottleneck
• Synthetic data generation for humanoids shows 20% improvement over real-world-only training in simulation benchmarks
• Physics-grounded tactile representations enable zero-shot sim-to-real transfer for specific dexterous tasks
• Cluttered scene manipulation remains extremely difficult, with real-world success rates around 50% even for state-of-the-art methods

That last point comes from a paper on extrinsic dexterity, which is the fancy term for using environmental contact to manipulate objects. According to arXiv, their method achieved about 50% success rate across 10 cluttered real-world scenes. Fifty percent! These are researchers who know what they're doing, working in controlled lab conditions, and they're hitting coin-flip odds. The paper frames this as a positive result because it beats prior methods by 25%, which it does, but let's not pretend this is anywhere close to deployment-ready.

There's one more paper worth mentioning, called BORA, which tries to bridge offline learning with online adaptation for dexterous manipulation. The approach uses human-in-the-loop correction during online learning to fix execution errors, which is basically admitting that the robot can't figure it out on its own. According to arXiv, this achieves a 33% absolute increase in success rate over pure imitation learning. That's substantial! But the fact that we need humans constantly intervening to make robot learning work tells you something about where we actually are versus where the press releases say we are.

I've been covering tech long enough to recognize the pattern. We're in the phase where the demos look magical, the benchmarks keep improving, and everyone's convinced the breakthrough is imminent. The young founders raising money on these ideas are smart and hardworking and genuinely believe they're close. Maybe they are! But the catastrophic forgetting problem isn't a bug to be patched. It's a fundamental challenge in how neural networks learn, and the fact that it shows up so clearly in real-world robot learning suggests we're still missing something important about how to build systems that actually generalize.

The researchers studying these problems are doing valuable work. They're being honest about limitations, publishing negative results, and building the foundations that might eventually lead somewhere useful. That's how science is supposed to work. What I'm less patient with is the gap between what the research actually shows and what gets repeated in investor decks and product announcements. A 50% success rate in a lab is not "robots are ready for your warehouse." A 20% improvement in simulation is not "humanoids will replace human workers by 2027."

But what do I know. If you want to argue about it, my email's on the about page.

James Chen · 8 hours ago · 4 min