You have two options when facing a Sim2Real gap: either bridge the gap by finding real-world applications for simulated scenarios or widen the gap to explore new and innovative ideas that might not be feasible in reality. A straightforward possibility is to introduce a novel incentive, instructing the robot not to engage in any undesirable behavior it’s currently exhibiting. Despite these rewards being implemented, they still feel like makeshift Band-Aids on the robotic system – unrefined and lacking a solid conceptual basis. They accumulate, often obscuring the distinct purpose of the report with an array of diverse descriptions. While it appears to provide adequate coverage, the outcome is unclear and becomes increasingly unreliable when combined with novel incentives.
Rather than embracing a less resilient option, it’s more prudent to investigate why certain simulations deviate from reality. As an organization, Agility has consistently focused on grasping the physiological underpinning of our actions. We meticulously crafted every aspect of our robotic system, from the actuators to the underlying software programming.
Our RL (Reinforcement Learning) method stands out distinctly from others. Let’s drive insights from our perceptions. We embarked on a six-month odyssey to uncover the mystery behind our simulated toes’ inability to replicate the exact same functions as their natural counterparts.
Markets are driven by numerous factors? Uncertainties persist due to oversimplifications within the collision geometry, discrepancies arising from vitality propagation through actuators and transmissions, and instability issues in constraint solving for our unique closed-chain kinematics, wherein connecting rods linked to toe plates and the tarsus exhibit complexity. We have been methodically addressing, rectifying, and eradicating these knowledge voids.
The online integration has yielded a substantial leap forward for our real-time software suite. Instead of piling on stacked-reward capabilities for trivial actions like “stop wiggling your foot” to “stand up straighter,” we’ve opted for a more thoughtful approach by focusing rewards around energy consumption and symmetry, which are not only simpler but also align with our core intuitions about how Digit should move.
By scrutinizing the reasons behind the simulation’s disparity, we’ve gained a deeper understanding of what drove our desire for Digit to take a deliberate approach in the first instance. Moreover, leveraging the rapid NVIDIA Isaac Sim, a cutting-edge reference tool built upon NVIDIA Omniverse for simulating and testing AI-driven robots, we have been able to uncover the implications of various physical attributes that will be essential in future iterations of Digit.
