Analysis
Why Robotics Isn't Solved Yet — And What Might Fix It
A YC robotics talk explains why every 'year of the robot' prediction has failed, and what two new research approaches are doing about it.
Quick Verdict
- The core takeaway
- Robotics hasn't been 'solved' by any recent hype cycle because physics simulation, action representation, tactile sensing, and hardware drift remain unsolved — but new work on giving policies memory and self-supervised reasoning is starting to chip away at the problem.
- Key tool featured
- Vision-Action Models (VLAs)
- Who this is for
- Robotics researchers, ML engineers, and founders evaluating how close general-purpose robots actually are to deployment.
Four Walls Nobody Has Solved
Robotics has promised a breakthrough year for a decade — AlphaGo, MuJoCo, Aloha, diffusion policy, vision-action models (VLAs). Each time, the field declared victory was near. The opening speaker argues we're still stuck behind four unsolved barriers, not one.
- Sim-to-real gap: world models used to simulate physics for robots don't actually respect real-world physics, especially with deformable objects.
- Action representation: conditioning a model on the action it's about to take, not just the current state, is still largely unsolved and data-hungry.
- Sensory-motor poverty: robots have a coin-sized force sensor per fingertip at best, versus a human's full-body network of touch, temperature, and friction sensing.
- Embodiment drift: actuators degrade, corrode, and shift in behavior over time, silently invalidating the teleoperation data a policy was trained on.
The practical upshot: teleoperation data collection, the backbone of most current training, is slow, finicky, and doesn't scale the way image or text data did for LLMs.
Giving Robots Memory
A Stanford PhD researcher who interned at Physical Intelligence presented work on a system called MAM (multiscale embodied memory). Most current robot policies, including PI0 and Groot, have no memory at all — every timestep is treated as brand new, with no sense of how long a task has been running.
Without memory, robots wash dishes forever or burn a grilled cheese because they have no internal clock. The fix splits memory into two tiers: a short-term visual memory (dense recent frames feeding a low-level action policy) and a long-term memory (a compressed text summary feeding a high-level planning policy).
- Short-term memory: built on a modified vision transformer with temporal attention, compressing frame history into the current token so inference stays fast.
Pros & Cons
Advantages
- Adding short- and long-term memory to VLA policies lets robots track task progress and stop repeating failed actions.
- Memory-based policies showed in-context adaptation — correcting a mistake after making it once, unlike memoryless baselines stuck in loops.
- Self-supervised reasoning bootstrapping (RB-CoT) improved success rates and out-of-distribution performance across manipulation, legged locomotion, and self-driving without needing hand-labeled reasoning data.
Limitations
- Textual memory representations may limit what the model can express compared to richer latent embeddings, and this hasn't been fully explored.
- Current long-term memory relies on supervised fine-tuning with human-annotated 'important information,' which is labor-intensive and not automatically optimal.
- Core barriers — sim-to-real physics mismatch, action-conditioned dynamics, tactile sensing, and actuator drift — remain unsolved and aren't addressed by either presented method.
Frequently asked
Why do robotics predictions of 'next year it'll be solved' keep failing?
According to the talk, each hype cycle (AlphaGo, MuJoCo, Aloha, diffusion policy, VLAs) solved one piece of the puzzle but left at least four fundamental barriers unaddressed: simulated physics not matching reality, poor action-conditioned world models, minimal tactile sensing compared to human skin, and hardware that drifts and degrades over time.
What is embodied memory in robot policies?
It's the idea of giving a robot policy a sense of history and time, split into short-term visual memory for fine motor tasks and long-term compressed text memory for tracking progress on multi-minute tasks like cooking a full meal.
What is RB-CoT in robotics?
Refine and Bootstrap Embodiment-specific Chain of Thought is a self-supervised training cycle that proposes and scores different types of reasoning traces (like object lists or gripper position) to find which kinds of reasoning actually improve action prediction for a given robot embodiment.