SCDP: Learning Humanoid Locomotion from Partial Observations via Mixed-Observation Distillation

The paper introduces Sensor-Conditioned Diffusion Policies (SCDP), a novel framework that enables robust humanoid locomotion using only onboard sensors by distilling privileged full-body knowledge through mixed-observation training and specialized denoising techniques, successfully achieving near-perfect simulation performance and real-world deployment on a G1 robot without explicit state estimation.

The Gist

Imagine you are trying to teach a robot to walk like a human. Usually, to do this, you give the robot a "superpower": a perfect, god-like view of the world. You tell it exactly where its feet are, how fast it's moving, and its exact orientation in space. This is like a robot having a GPS, a speedometer, and a gyroscope all fused into one perfect brain.

The Problem:
In the real world, robots don't have these superpowers. They only have "onboard sensors"—basically, they can feel their own joints moving and sense gravity, but they can't "see" their own speed or exact position without expensive external cameras. If you take away the "superpower" view from existing robot controllers, they usually fall over immediately. It's like trying to drive a car with your eyes closed, relying only on the feeling of the steering wheel.

The Solution: SCDP (Sensor-Conditioned Diffusion Policies)
The researchers at UCL created a new method called SCDP. Think of it as a "magic trick" for teaching robots to walk using only their internal senses, without needing that perfect external view.

Here is how they did it, broken down into simple analogies:

1. The "Secret Teacher" Game (Mixed-Observation Distillation)

Imagine you are learning to play a complex video game.

• The Old Way: You practice with a cheat sheet that shows you the enemy's exact location and speed. When you try to play without the cheat sheet, you fail because you never learned to guess where the enemy is.
• The SCDP Way: The robot plays a game where it only sees its own hands and feet (the sensors). However, the "teacher" (the training algorithm) secretly grades it based on a cheat sheet that does show the enemy's location.
• The Result: The robot is forced to look at its limited view (hands/feet) and figure out, "If my left foot is moving this way and I feel this much wind, I must be moving at 2 meters per second." It learns to infer the hidden information (speed and position) just by looking at the clues it does have. It becomes a detective of its own motion.

2. The "Blindfolded Sprint" (Restricted Denoising)

One specific clue robots usually rely on is "how fast am I going?" (velocity). But real robots are bad at measuring their own speed accurately.

• The Trick: During training, the researchers told the robot: "I will tell you where you should be going, but I will hide the speed number from your input."
• The Analogy: Imagine you are running a race blindfolded. Your coach yells, "Go faster!" but doesn't tell you your current speed. You have to feel the wind and your leg muscles to guess how fast you are running and adjust accordingly.
• The Outcome: The robot learns to estimate its own speed purely from the "feel" of its movement, making it robust even if its sensors are noisy or imperfect.

3. The "Time-Traveling Memory" (Context Distribution Alignment)

When teaching the robot, the researchers had to be careful not to trick it.

• The Problem: If you train a robot using "noisy" (messy) data but then let it run on "clean" data in the real world, it gets confused. It's like practicing basketball with a slippery ball, then trying to play with a dry one.
• The Fix: They made sure the robot practiced in a way that perfectly matched the real world. They also gave the robot a "memory window" (context) where it could look back at its recent history to understand the present.
• The Analogy: Instead of looking at a single frozen frame of a movie, the robot watches the last few seconds of the film to understand the plot. This helps it predict what will happen next, even if it can't see the whole future.

The Results: From Simulation to Real Life

The team tested this on a Unitree G1, a real humanoid robot that looks like a futuristic human.

• In the Computer: The robot learned to walk, turn, and recover from pushes with nearly 100% success, matching robots that had "superpowers" (perfect sensors).
• In the Real World: They put the robot in a real room. Without any external cameras or motion-capture suits, the robot walked around at 50 times a second (50 Hz), reacting to pushes and following commands smoothly.

Why This Matters

Before this, if you wanted a robot to walk like a human, you needed a lab full of expensive cameras to track its every move. SCDP changes the game. It proves that a robot can learn to be a "self-aware" walker, inferring its own speed and position just by feeling its own body, much like a human does when they close their eyes and walk.

In a nutshell: They taught a robot to "feel" its way through the world, turning a blind, stumbling machine into a confident, self-aware walker using a clever training game where it had to guess the hidden rules of physics.

Technical Summary

Here is a detailed technical summary of the paper "SCDP: Learning Humanoid Locomotion from Partial Observations via Mixed-Observation Distillation."

1. Problem Statement

Current humanoid locomotion controllers, particularly those based on Diffusion Models and Offline Distillation, rely heavily on privileged full-body state information (e.g., global position, orientation, and base velocity) during deployment.

• The Challenge: In real-world scenarios, these privileged states are not directly observable. Obtaining them requires complex, often unreliable state estimation pipelines (e.g., motion capture, SLAM, or sensor fusion).
• The Consequence: Removing privileged inputs causes a catastrophic performance drop in existing methods, as they cannot infer hidden global states from partial onboard observations (a Partially Observable Markov Decision Process, or POMDP).
• The Goal: Develop a policy that learns complex whole-body behaviors from offline datasets using only onboard proprioceptive sensors (joint positions, velocities, IMU data) without explicit state estimation.

2. Methodology: Sensor-Conditioned Diffusion Policies (SCDP)

The authors propose SCDP, a framework that distills an expert Reinforcement Learning (RL) policy into a diffusion-based policy capable of operating under partial observability. The core innovation is Mixed-Observation Distillation, which decouples sensing from supervision.

Core Components:

1. Mixed-Observation Training:

   • Conditioning (Input): The diffusion model is conditioned only on historical onboard sensor data ($o_t$) and task commands.
   • Supervision (Target): The model is supervised to predict privileged full-body state-action trajectories ($s_t, a_t$), including global positions and velocities that are unavailable at runtime.
   • Mechanism: This asymmetry forces the model to learn an implicit internal representation of global body dynamics, effectively "hallucinating" the missing state information from the sensor history.

2. Restricted Denoising:

   • A specific challenge is controlling velocity without direct velocity feedback (as onboard velocity estimates are noisy).
   • Technique: During the denoising process, the model is fed trajectories excluding the pelvis linear velocity ($v_{pelvis}$). However, the supervision target includes the true velocity.
   • Effect: This forces the model to infer velocity purely from the context of sensor history and task commands, rather than relying on noisy direct feedback.

3. Context Distribution Alignment:

   • Standard diffusion training often uses noisy states ($s_{noised}$) derived from noisy actions, creating a distribution mismatch with the clean states observed at inference.
   • Technique: The training context is constructed using $(s_{noised}, a_{noised})$ pairs generated by executing noisy actions. This aligns the training distribution with the inference conditions, reducing distribution shift.

4. Context-Aware Attention Masking:

   • Unlike prior works that use strictly causal masks, SCDP enables bidirectional attention within the context window (past observations).
   • Effect: This allows the model to aggregate historical information more effectively to infer latent dynamics, while maintaining causal constraints only for the future prediction horizon.

5. Architecture:

   • Expert: A Multi-Motion Tracking Policy (MMP) trained via PPO on diverse walking motions (AMASS dataset).
   • Student: A 6-layer Transformer-based diffusion policy. It processes sensor history, commands, and past actions to denoise future state-action trajectories.

3. Key Contributions

• Mixed-Observation Distillation: A novel training formulation that enables diffusion policies to learn global dynamics from partial sensor data by predicting privileged states.
• Velocity Control without Feedback: Introduction of Restricted Denoising, enabling robust velocity control without relying on direct, unreliable velocity sensors.
• Robust Real-World Deployment: Successful deployment on a Unitree G1 humanoid robot at 50 Hz, demonstrating robust locomotion without external motion capture or state estimation.
• Systematic Ablations: Comprehensive analysis of data collection strategies (noise/pushing), context length, and architectural components, identifying that mixed-observation training is the foundational requirement for success.

4. Experimental Results

Simulation Performance (IsaacLab):

• Velocity Control: Achieved 99–100% success in perturbation recovery, navigation, and joystick control tasks.
• Comparison: SCDP (using only onboard sensors) performed comparably to privileged baselines (which have access to full state) and significantly outperformed standard Behavior Cloning (BC) or context-only diffusion policies (which failed with <10% success in some tasks).
• Motion Tracking: Achieved 93% success in tracking reference motions from the AMASS test set, with a global position error (MPJPE-G) of ~288mm, significantly better than BC baselines.

Real-World Deployment (Unitree G1):

• Frequency: The policy runs at 50 Hz on a remote workstation (RTX 5090) with ONNX Runtime.
• Performance: Demonstrated robust walking, turning, and recovery from perturbations in a real-world setting without external sensing.
• Velocity Tracking: Showed smoother linear velocity tracking with reduced oscillations compared to the "BeyondMimic" baseline, though with a slight lag (~250–500ms) during direction changes.

Ablation Insights:

• Mixed-Observation Training: Essential. Removing it caused success rates to drop to ~1.4%.
• Representation Alignment: Surprisingly, explicitly aligning sensor and privileged state embeddings hurt performance; the model learns better implicitly.
• Context Length: Shorter contexts (4–8 steps) performed better than long histories (16 steps), suggesting that longer histories amplify compounding errors in POMDP settings.

5. Significance

This work addresses a critical bottleneck in humanoid robotics: the gap between simulation-trained policies (which assume perfect state knowledge) and real-world deployment (where state estimation is difficult).

• Eliminates State Estimation: SCDP proves that complex locomotion can be learned without explicit state estimation pipelines, making deployment more robust and scalable.
• Generative Control: It demonstrates that diffusion models, when properly trained with mixed observations, can effectively solve POMDPs for high-dimensional robotic control.
• Sim-to-Real Transfer: The successful deployment on the Unitree G1 validates the framework's ability to transfer from simulation to reality without fine-tuning on privileged data.

In summary, SCDP provides a robust, generalizable framework for training humanoid controllers that rely solely on onboard sensors, bridging the gap between high-performance simulation training and practical, real-world robot deployment.