🤖 machine learning

PGTT: Phase-Guided Terrain Traversal for Perceptive Legged Locomotion

The paper introduces PGTT, a perception-aware deep reinforcement learning framework that employs phase-guided reward shaping to enable robust, morphology-agnostic legged locomotion across diverse terrains without relying on restrictive gait priors or blind control strategies.

Original authors: Alexandros Ntagkas, Chairi Kiourt, Konstantinos Chatzilygeroudis

Published 2026-07-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alexandros Ntagkas, Chairi Kiourt, Konstantinos Chatzilygeroudis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where robots don't just shuffle along on wheels, but hop, skip, and jump over messy, uneven ground just like a dog or a human. This is the exciting frontier of legged robotics. For a long time, making these robots move was like teaching a toddler to walk: you had to give them very strict rules, like "lift your left foot exactly this high, then your right foot exactly that high." This worked, but it made the robots stiff and confused when they hit something unexpected. Then, scientists discovered Reinforcement Learning (RL), a method where robots learn by trial and error, kind of like a video game character learning to jump over gaps by falling a thousand times and getting better each time. However, many of these "learning" robots were either "blind" (they couldn't see the ground ahead) or they were still stuck with those old, rigid rules that made them clumsy on new types of terrain. The big question researchers are trying to solve is: How do we teach a robot to be both smart enough to see the ground and flexible enough to adapt its steps without being told exactly how to move every single joint?

Enter PGTT (Phase-Guided Terrain Traversal), a new approach that tries to find the perfect middle ground. Think of the old "blind" robots as people walking with their eyes closed, hoping they don't trip. The old "rule-bound" robots are like dancers who must follow a choreographer's exact script; if the floor changes, they can't improvise. PGTT is like teaching a robot to dance with a rhythm but letting it improvise the steps.

The researchers gave the robot a pair of "smart eyes" (a laser scanner) that builds a quick, 3D map of the ground right in front of it. Instead of forcing the robot to hit specific foot targets, they taught it a simple rhythm: "Your legs should swing in a wave." But here's the magic trick: they didn't force the robot's joints to follow a strict path. Instead, they gave the robot a scorecard (a reward system). If the robot swings its leg too low and hits a rock, it gets a bad score. If it swings high enough to clear the rock, it gets a good score. The robot learns to figure out how to move its own muscles to get that good score, adapting its step height automatically based on what it sees.

The paper shows that this method works incredibly well in computer simulations. When tested on tricky, stair-like terrain and scattered obstacles, the PGTT robot was much more successful at staying upright than other top methods. In fact, it survived 7.5% more pushy disturbances (like someone shoving it) and cleared 9% more obstacles than the next-best robot. It didn't just survive; it kept moving at the same speed as the others.

The team also tried this on a real robot, the Unitree Go2 (a four-legged robot that looks a bit like a dog). They used the same "brain" they trained in the computer, without changing any settings, and it successfully climbed real stairs and hopped over real obstacles. They even tested it on a different robot, the ANYmal-C, and it worked there too, suggesting this "rhythm-without-rules" idea might work for all kinds of legged machines.

However, the paper is careful to note that this isn't a magic bullet for every situation yet. The real-world tests were done at a modest walking speed of 0.4 meters per second because the robot's laser scanner isn't fast enough to update the map quickly at higher speeds. Also, while the robot is great at not falling, the researchers haven't measured yet if it's using energy efficiently. But overall, PGTT suggests that by giving robots a simple rhythm to follow but letting them figure out the details themselves, we can build machines that are robust, adaptable, and ready to tackle the messy, unpredictable real world.

Technical Summary: Phase-Guided Terrain Traversal (PGTT)

Problem Statement
State-of-the-art perceptive Reinforcement Learning (RL) controllers for legged robots face a dichotomy in design. Existing approaches typically either (i) impose oscillator- or Inverse Kinematics (IK)-based gait priors that constrain the action space, bias policy optimization, and limit adaptability across different robot morphologies, or (ii) operate "blindly" (relying solely on proprioception), rendering them unable to anticipate hind-leg terrain and making them brittle to observation noise. While visual or range sensing has been integrated into control loops, challenges remain regarding limited fields of view, sensor noise, and the need for temporal memory or brittle global pose estimation. The core challenge is to achieve robust, terrain-adaptive locomotion that anticipates obstacles without introducing excessive inductive bias or coupling the policy to specific robot kinematics.

Methodology
The authors propose Phase-Guided Terrain Traversal (PGTT), a perception-aware deep-RL approach that enforces gait structure through reward shaping rather than action-space constraints.

Perception: PGTT utilizes a compact, robot-centric heightmap derived online from LiDAR elevation mapping. This local representation captures terrain geometry relevant to all legs without requiring global consistency or pose estimation, feeding directly into the policy network.
Gait Encoding: Instead of using oscillators or IK to prescribe joint targets, PGTT encodes a per-leg phase variable $\phi_{i,t}$ as a cubic Hermite spline. This spline defines a smooth, terrain-adaptive swing trajectory where the swing apex adapts to local heightmap statistics (specifically, the maximum and minimum terrain heights around the leg).
Reward Shaping: The phase prior is enforced exclusively through the reward function, not the action space. The policy acts directly in joint space (outputting desired joint angles relative to a stand-still pose), preserving an unconstrained action space. The reward includes:
- A phase-guided term encouraging feet to follow the Hermite-spline trajectory.
- A swing-phase contact penalty to prevent premature ground contact during the swing phase.
- Standard task rewards (velocity tracking, orientation, joint limits).
Learning Architecture: The system employs an asymmetric actor-critic architecture trained with Proximal Policy Optimization (PPO) in MuJoCo (specifically the GPU-accelerated MJX branch). The actor conditions only on partial observations (proprioception + heightmap), while the critic utilizes privileged state information (including true linear velocity) for value estimation.
Training Strategy: Training occurs in procedurally generated stair-like environments using the Wave Function Collapse (WFC) algorithm. A curriculum learning approach progressively increases terrain difficulty (step height and complexity). Extensive domain randomization (sensor noise, robot physical parameters, and friction) is applied to bridge the Sim-to-Real gap.

Key Contributions

Terrain-Adaptive, Phase-Guided Reward: A novel reward formulation that encodes Hermite-spline swing trajectories driven by local heightmap statistics and penalizes swing-phase contacts. This approach reduces inductive bias compared to oscillator/IK-conditioned methods and facilitates morphology-agnostic deployment.
Compact Perception-to-Policy Design: A system that feeds a robot-centric heightmap directly to the policy, avoiding global pose assumptions while capturing necessary terrain geometry for stepping.
Extensive Evaluation and Transfer: Comprehensive simulation results on stair-like and discrete-obstacle terrains demonstrating higher success rates than baselines. The work includes Sim2Real transfer on a Unitree Go2 and preliminary cross-platform results on an ANYmal C using identical hyperparameters.
Accessible Training Stack: A lightweight training pipeline using MuJoCo/MJX that achieves high throughput on a single consumer GPU, offering an alternative to heavier Isaac Gym-based pipelines.

Results

Simulation Performance: PGTT achieved the highest success rate among evaluated baselines (MassLoco and Wild) under push disturbances (median +7.5% over the next-best) and on discrete obstacles (+9%). Velocity tracking performance remained comparable to the strongest baselines.
Convergence: PGTT demonstrated faster learning convergence than an end-to-end, non-prior RL baseline and significantly better sample efficiency than MassLoco.
Real-World Deployment: The policy was successfully deployed on a Unitree Go2 using a real-time LiDAR elevation-to-heightmap pipeline, demonstrating robust stair and obstacle traversal. Preliminary results on an ANYmal C showed successful walking behaviors without any hyperparameter retuning.
Ablation: Removing the foot phase reward reduced the success rate on stairs from 83.4% to 75.5%, confirming it as the primary driver of performance. The foot contact penalty provided a smaller but necessary refinement for edge-case stability.

Significance and Claims
The paper claims that PGTT provides early evidence that terrain-adaptive, phase-guided reward shaping can transfer across different robot platforms without platform-specific policy priors or extensive retuning. By retaining the benefits of rhythmic structure through reward shaping while keeping the action space unconstrained, the method reduces inductive bias and supports morphology-agnostic deployment. The authors position PGTT as an accessible foundation for advancing agile, robust legged locomotion, lowering the barrier to entry for researchers by enabling training on affordable hardware (single consumer GPU) while maintaining high performance.

Limitations
The authors acknowledge that the current implementation relies on a lower-specification L1 LiDAR (21,600 points/s), which restricts experiments to a maximum speed of 0.4 m/s due to sparser measurements at higher speeds. Additionally, the energy efficiency of PGTT in real-world deployments has not yet been evaluated.

Technical Summary: Phase-Guided Terrain Traversal (PGTT)

More like this