Furkan Gözükara
patreon
Real Steel Became a Reality - Full AI Robots Boxing Tournament - With English Subtitles - So How These Robots Working Actually in Full Details
🕑 Added 2025-05-27 20:42:17 +0000 UTCThis video is from Live: CMG World Robot Competition – Mecha fighting series with English Subtitles. China started to rule Open Source generative AI and now they are moving to robotics earlier from everyone else. Who would have guessed just after 14 years of Real Steel movie, we will have real Real Steel. And this is only the beginning. It will only get better very likely and fast. These are commercial level robots so they are cheap and weak. Still they perform excellent.
4K HD Video Link : https://youtu.be/fw2yezpn_bo
4K HD Video Link
How These Robots Working?
It's a hybrid system, with significant AI handling the low-level execution and complex movements, while humans provide high-level strategic commands and possibly trigger specific actions.
1. AI-Driven Aspects (Significant AI Usage):
Movement & Balance:
Bipedal Locomotion: The robots walk, sidestep, and maintain balance using sophisticated algorithms. This is a huge AI challenge for humanoid robots.
Getting Up After Falls (Adaptive Recovery): A lot of emphasis is placed on how quickly and "coolly" the robots get back up. The commentators mention this is trained using large amounts of human motion capture data (specifically citing professional boxers' movements for getting up). The robots don't just play one animation; they seem to adapt their recovery to the situation (as noted around 9:50 when they say the posture and angle are different each time). This indicates AI-driven dynamic recovery.
Dynamic Stability: The robots have to constantly adjust to maintain balance during rapid movements and after impacts. This involves sensor fusion (IMUs, vision, possibly tactile) and AI control loops.
Motion Control & Execution of Actions:
Pre-trained Complex Moves/Combos: Specific named attacks (like "胭脂紅 - Rouge Red") are described as very coherent, pre-trained sets of movements. The entire sequence, including footwork and arm movements, is a coordinated "motion control algorithm." The human operator likely triggers "Combo A" rather than controlling each limb.
Fluency of Skills: Professor Li Gaofeng (the expert commentator) highlights the "fluency of skills" (流暢度). He explains that robots learn these skills (again, from human boxer data) and the AI's job is to map these human motions effectively to the robot's different kinematics and dynamics.
Taunts/Personality: Programmed taunts (like patting the thigh) are designed actions, likely triggered.
Sensory Processing & Reaction (Limited):
Vision System: Mentioned that they have an "autonomous vision system to collect various information" (around 3:46). This would be crucial for identifying the opponent, distance, etc.
Impact/Force/Tilt Sensors: The robots have sensors to detect when they are hit and their degree of tilt (around 3:00-3:05), which feeds into their balance and reaction algorithms. They don't have "pain" sensors in the human sense; this is a deliberate design choice to make them fearless. They do have force/tactile sensors.
Safety Protocols:
Robots are programmed to stop their current action immediately if they fall (around 10:00-10:20). This is a safety feature to prevent self-damage or damage to the arena.
2. Human-Controlled Aspects (Strategic & Triggering):
High-Level Strategy & Decision-Making:
When to Attack/Defend/Taunt: The human operators ("操縱手 - manipulators/operators") are almost certainly deciding the overall fight strategy, choosing when to be aggressive, defensive, or use a taunt.
Initiating Actions/Combos: While the AI executes the complex motor control for a combo, the human operator likely initiates these pre-programmed sequences. The commentary mentions operators choosing to "use one successful kicking action to get three points" (8:07), indicating strategic selection of moves.
Adapting to Opponent: The expert, Li Gaofeng, emphasizes "human-robot collaboration" (人機協同) where the human operator makes decisions based on the robot's current state and the opponent's actions, then issues commands.
Potential Fine-Tuning of Actions:
Strike Intensity/Risk Management: Around 7:33, there's a crucial discussion. When a robot lands a heavy punch, its motors experience a surge in current, which can be risky. The commentator says, "our contestants (operators) should control this to some extent." This strongly suggests that human operators might have some control over the intensity or timing of a strike, not just triggering it. This could be to manage power consumption or prevent motor burnout.
"Piloting" with Abstracted Commands:
The commentators distinguish these robots from simple remote-controlled toys where one joystick controls one limb. Here, human commands are likely more abstract, e.g., "move forward," "execute combo X," "high punch." The AI then translates this into the necessary joint movements using its "large models" and "motion control algorithms" (as discussed around 3:26-3:37).
How Much AI vs. Human Control?
AI is dominant in:
Executing complex physical movements (walking, balancing, combos, getting up).
Low-level motor control and coordination.
Interpreting sensor data for balance and impact.
Running safety protocols.
Humans are dominant in:
Overall fight strategy and high-level decision-making.
Timing and selection of specific attacks or defenses (triggering pre-programmed AI routines).
Potentially modulating the intensity or risk of certain actions.
In essence:
Think of it like a video game character. The AI makes the character walk smoothly, perform complex attack animations, and react to physics. The human player decides when to walk, which attack to use, and the overall strategy. However, in this robot boxing, the "animations" (robot movements) are incredibly sophisticated and adaptive due to extensive AI training and real-time algorithmic control. The human operators are skilled "pilots" giving strategic instructions to a very capable AI-driven physical avatar.
The expert commentary really drives home that the fluency and recovery capabilities are a result of AI learning from vast amounts of human data, while the strategic layer and the moment-to-moment decision-making (what move to use now) involve significant human input.
