93 Hours of Training Just to Walk? Bitbots Drops First Teaser

When "Learning to Walk" Becomes a Marathon

A Reddit post has sparked plenty of discussion: "My robot had to train for a full 93 hours before it could walk," accompanied by Bitbots' first teaser. It sounds like a complaint, but it precisely captures the real cost of legged robot training today — reinforcement learning-driven locomotion control is still a time-devouring endeavor.

What does 93 hours mean? If you train 8 hours a day nonstop, that's nearly 12 days. On a real physical robot, it means a long process of falling, adjusting, and falling again. The teaser doesn't show off flashy backflips — instead, it uses "clumsy first steps" as a selling point. That kind of honesty is itself worth pondering.

Why Is Learning to Walk So Hard?

The "Time Black Hole" of Simulation Training

Legged robot gait control is essentially a high-dimensional continuous control problem. Joint angles, ground friction, center-of-mass shifts… any tiny change in a single parameter can turn a robot from "walking steadily" into "face-planting instantly."

The mainstream approach is to train policies with reinforcement learning in simulation, then transfer them to real hardware. But those "93 hours" in simulation aren't just idle runtime — they require massive parallel environments sampling simultaneously, which is a hard drain on compute. What's worse is the sim-to-real gap: no matter how stable the gait is in simulation, once the real robot encounters carpet wrinkles or lighting changes, the policy may fail outright.

Reinforcement Learning (RL) is currently the dominant paradigm for legged robot locomotion control. Its basic framework is: an agent continuously tries actions in an environment, adjusts its policy based on reward signals received, with the goal of maximizing long-term cumulative reward. In the robot walking scenario, the reward is typically designed as "move forward quickly without falling," and the agent is a neural network policy that outputs target angles or torques for each joint.

The key difference between RL and traditional control methods (such as MPC model predictive control) is that it doesn't require precise physical modeling — instead, it "learns" a policy through massive trial and error. The cost is extremely low sample efficiency: a policy that can walk often requires tens of millions or even hundreds of millions of simulation interactions. This is why "93 hours" sounds exaggerated but is actually normal in the industry.

To accelerate training, the industry widely adopts large-scale parallel simulation (such as Isaac Gym, MuJoCo MJX, and other GPU-accelerated platforms), throwing thousands of robot instances into virtual environments simultaneously and using GPUs for parallel sampling. This is what compresses training that would otherwise take months into tens of hours.

The sim-to-real gap is one of the core obstacles to deploying legged robots. No matter how refined the simulation environment is, it cannot fully reproduce real-world physical details: motor response delays and friction, tiny ground deformations, sensor noise, lighting and texture variations… These differences cause policies that perform perfectly in simulation to become "fish out of water" on real hardware.

The industry commonly uses three categories of countermeasures: first, Domain Randomization — randomly perturbing parameters like friction, mass, and latency during training so the policy learns to adapt to various changes; second, System Identification — using real robot data to calibrate simulation parameters and narrow the gap; third, Real-world Fine-tuning — continuing training on the simulation policy with a small amount of real robot data.

This is also why "93 hours of simulation training" doesn't equal a mature product — going from simulation to stable real-world performance often requires several times that amount of debugging and validation work.

Behind the 93 Hours Is Policy Iteration

The "93 hours" mentioned in the teaser most likely refers to simulation training time, not real robot practice. This number isn't exaggerated in the industry — training quadruped gaits often requires tens of millions or even hundreds of millions of simulation steps. 93 hours is more like a carefully chosen "perceptible" number: it shows the barrier is high while also implying "we actually pulled it off."

Quadruped robots are far more forgiving in locomotion control than bipedal humanoids: quadrupeds are inherently statically stable — even if one or two legs are in poor condition, the remaining legs can maintain balance. This is why consumer-grade robots generally start with quadrupeds — Unitree's Go series, Xiaomi's CyberDog, and Boston Dynamics' Spot are all quadruped forms.

Quadruped gait training typically proceeds in several stages: first basic standing and balance, then periodic gaits like trot (diagonal pairs), and finally more complex bound (jumping-style), pace (same-side), as well as stair climbing and resistance to external disturbances. Each stage corresponds to different reward designs and curriculum learning strategies.

"Learning to walk" typically means mastering a basic trot gait and being able to move stably. From trot to truly handling complex terrain, there's still a great deal of engineering work in between — which is why the "93 hours" in the teaser should be understood as a starting point, not an endpoint.

What Does the Bitbots Teaser Reveal?

Judging from the teaser screenshots, Bitbots is taking a miniaturized, lightweight approach. This stands in sharp contrast to Boston Dynamics' hydraulic behemoths, and is closer to the product logic of Xiaomi's CyberDog and Unitree's Go series: using consumer-grade hardware + reinforcement learning to bring legged robots into a mass-producible price range.

The teaser deliberately emphasizes "trained 93 hours just to learn to walk," which may have two intentions: first, to establish technical credibility by implying the team put genuine effort into locomotion control; second, to lower expectations and avoid unfair comparisons with mature products. For a new product not yet validated at scale, this narrative strategy is quite pragmatic.

Training Time Is Becoming a New Selling Point

Interestingly, "how long it trained" is evolving from an engineering detail into marketing language. In the past, manufacturers loved to say "how many TOPS of compute we use" or "how many sensors we have onboard." Now they're starting to say "how many hours it trained" or "how many times it fell."

Behind this is a shift in industry narrative: hardware specs are becoming increasingly homogenized — training methods and data accumulation are the real moat. Whether 93 hours is enough to walk resonates more with the developer community than what motors are used — after all, those folks on Reddit have stayed up all night staring at loss curves running their own reinforcement learning experiments.

Of course, training time doesn't equal product maturity. Learning to walk in 93 hours only means basic locomotion capability has taken shape — there's still a long way to go before stable stair climbing, disturbance rejection, and long battery life. After the teaser, the real test is mass-production consistency and real-world robustness.

Summary

Bitbots uses the number "93 hours" to make the barrier of legged robot training tangible. It reminds us that making robots walk was never magic — it's an accumulation of compute, algorithms, and time. Whether this product can gain a foothold in the crowded consumer robot market remains to be seen after the teaser.

For developers, what's worth paying attention to isn't the number 93 itself, but whether the training framework behind it is open-source and whether the sim-to-real approach is reproducible. That's what the Reddit community will really be asking about.