PaperlayerView exact revision on arXiv

Source-linked AI summary

SayTap: Language to Quadrupedal Locomotion

Yujin Tang, Wenhao Yu, Jie Tan, Heiga Zen, Aleksandra Faust, Tatsuya Harada

arXiv:2306.07580v3cs.RO

TL;DR

LLMs can plan at a high level, but interpreting low-level robot commands remains challenging. SayTap uses foot-contact patterns to connect natural-language commands to quadruped locomotion control. On 30 tasks, it reports higher contact-pattern prediction success and more tasks solved than its baselines.

  • Problem

    LLMs can perform high-level planning, but interpreting low-level commands such as joint targets or motor torques remains challenging.

    Source: §1 · p. 1

  • Method

    The approach uses an LLM to translate natural-language instructions into foot-contact patterns and a DRL controller to realize the desired patterns.

    Source: §1 · pp. 1–2

  • Results

    On 30 tasks, the approach achieves 50% higher success rate in predicting correct contact patterns and solves 10 more tasks than the baselines.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors conclude that contact patterns are a more straightforward and flexible interface than other design choices.

    Source: §5 · p. 9

    Designing a random pattern generator that produces feasible training patterns requires domain knowledge and trial-and-error, while training on many gaits trades off sample balancing and data efficiency.

    Source: Limitations and Future Work · p. 9

Abstract

from arXiv · show

Large language models (LLMs) have demonstrated the potential to perform high-level planning. Yet, it remains a challenge for LLMs to comprehend low-level commands, such as joint angle targets or motor torques. This paper proposes an approach to use foot contact patterns as an interface that bridges human commands in natural language and a locomotion controller that outputs these low-level commands. This results in an interactive system for quadrupedal robots that allows the users to craft diverse locomotion behaviors flexibly. We contribute an LLM prompt design, a reward function, and a method to expose the controller to the feasible distribution of contact patterns. The results are a controller capable of achieving diverse locomotion patterns that can be transferred to real robot hardware. Compared with other design choices, the proposed approach enjoys more than 50% success rate in predicting the correct contact patterns and can solve 10 more tasks out of a total of 30 tasks. Our project site is: https://saytap.github.io.

1 Introduction

LLMs can plan at a high level, but low-level robot commands remain difficult to interpret, limiting the expressiveness of API-based control. SayTap uses foot-contact patterns to connect natural-language instructions with quadruped control, and reports improved benchmark performance on 30 tasks.

1 Introduction

  • LLMs have difficulty interpreting low-level commands such as joint targets and motor torques, while high-level APIs can limit expressive capabilities.Natural language is a key communication channel for human-robot interaction.Source: §1 · p. 1
  • Foot-contact patterns bridge natural-language instructions and low-level commands, giving users a flexible interface for crafting locomotion behaviors.The paper presents an LLM pipeline for generating contact patterns and a DRL controller for realizing them.Source: §1 · pp. 1–2
  • Figure 1 pairs direct and vague commands with LLM-translated desired contact patterns and the robot’s realized patterns.The figure includes a pattern requesting all feet on the ground and the robot to stand still.Source: Figure 1 · p. 2
  • Contact timing often governs quadruped locomotion, so patterns of feet making and breaking ground contact serve as a compact interface for diverse behaviors.The controller is trained with a reward focused on contact timing and exposed to a suitable distribution of contact patterns; it can transfer to real hardware.Source: §1 · p. 2
  • On 30 tasks, the approach achieves 50% higher success rate in predicting correct contact patterns and solves 10 more tasks than the baselines.The baselines use discrete gaits and sinusoidal functions as interfaces.Source: §1 · p. 2

2 Related Work

Prior language-to-robot approaches use structured inputs, NLP-assisted planning, or language-conditioned policies, while locomotion controllers have used MPC and several learning-based strategies. This work positions contact patterns and a contact-timing-focused reward as flexible alternatives for quadruped control.

2.1 Language to robot control

  • Earlier language-to-robot methods use structured templates or parse-tree constraints, while language-conditioned policies for unstructured instructions often require substantial language-labeled demonstration data.The paper identifies contact patterns as an intermediate interface that does not require laborious design efforts.Source: §2.1 · p. 3
  • LLM-based robot systems connect language to commands through high-level plans, primitive skills, or trajectories; this work uses foot-contact patterns as an intermediate interface.The authors describe contact patterns as a natural and flexible interface for quadruped locomotion.Source: §2.1 · p. 3

2.2 Locomotion controller for legged robots

  • MPC can follow desired contact patterns but often requires careful reference-motion design and costly replanning; learning-based methods use varied gait-specific strategies.This work explores a DRL reward focused only on contact timing, which the authors characterize as more flexible for generating diverse locomotion behaviors.Source: §2.2 · p. 3

3 Method

SayTap uses foot-contact patterns as an intermediate interface: an LLM translates commands into patterns, and a learned controller realizes them alongside task goals. Training samples varied patterns, while the controller is trained to track task commands and realize the desired contacts.

3 Method

  • The system uses foot-contact patterns between natural-language commands and the locomotion controller, which also receives sensory and task inputs and outputs joint positions.Patterns are generated randomly during training and translated from commands by an LLM at test time.Source: §3 · pp. 3–4

3.1 Language to Foot Contact Patterns

  • The prompt combines a general translation instruction, gait definitions, a formatted output with discretized velocities, and examples to guide command-to-pattern generation.The prompt’s gait descriptions are not exhaustive or fully accurate, but are reported to provide enough information for the LLM to follow gait rules.Source: §3.1 · p. 5

3.2.1 Problem Formation

  • Locomotion is formulated as an MDP whose policy must accomplish the task while realizing the desired foot-contact pattern included in the state.The objective maximizes expected accumulated reward.Source: §3.2.1 · p. 5

3.2.2 Random Pattern Generator

  • The random generator samples cycle lengths and per-foot contact ratios, then applies scaling and phase shifts to create templates for five gait types.The types are bound, trot, pace, stand still, and stand on three legs.Source: §3.2.2 · p. 5

3.2.3 Locomotion Controller

  • A feed-forward policy maps robot state, velocity commands, prior output, and desired contacts to joint positions; a mirrored double pass improves movement naturalness and helps shrink the sim-to-real gap.The policy uses three hidden layers; its input and output dimensions are 65 and 12, respectively.Source: §3.2.3 · pp. 5–6

3.2.4 Task and Training Setups

  • Training rewards task tracking and matching desired foot-strike timing, while action-rate regularization smooths real-world movement and base-stability rewards help shrink the sim-to-real gap.The authors also report that balancing gait-type ratios during training helps, despite no heavy engineering being required for the added contact inputs.Source: §3.2.4 · p. 6

4 Experiments

Across the interface tests, contact-pattern prompts achieved about 50% higher accuracy than the baselines. The controller tracked velocity and realized contact patterns in simulation, transferred to hardware without fine-tuning, and produced mostly expected motions for vague instructions.

4 Experiments

  • GPT-4 was used throughout; each command was submitted five times at temperature 0.5, and only top-1 responses were compared.Source: §4 · p. 6, §4.1 · p. 6

4.1 Is Foot Contact Pattern a Good Interface?

  • The interface test compared contact patterns with five discrete gait types and a four-foot sinusoidal parameterization across basic and generalization or composition commands.Commands 1–20 were basic; commands 21–25 tested generalization and pattern composition.Source: §4.1 · p. 6
  • ∼50% higher accuracy than the baselines was achieved by the proposed approach across commands and test trials.Figure 5 reports aggregate accuracy alongside individual command results.Source: §4.1 · p. 6, Figure 5 · p. 8
  • Baseline prompts modified the proposed prompt while retaining the gait-definition block; black explanatory text was excluded from LLM input.Source: Figure 4 · p. 7

4.2 Can we learn to accomplish the main task and realize the contact pattern?

  • In simulation, the controller tracked specified velocity and produced contact patterns close to the desired ones, then transferred to the physical robot without fine-tuning.Figure 6 plots actual heading-direction velocity, desired contacts, and realized contacts across three trials.Source: §4.2 · p. 8, Figure 6 · p. 8

4.3 Does the proposed approach work with unstructured/vague instructions?

  • For vague instructions that omitted explicit actions or gait choices, observed motions mostly matched expectations in subjective evaluation.Examples included jumping when excited and limping when told the robot had a limping leg.Source: §4.3 · p. 8, Table 2 · p. 8

5 Conclusions

The system uses desired foot-contact patterns to flexibly connect natural-language commands with quadrupedal locomotion, including direct and implicit instructions in simulation and on physical robots. Its training approach has limits in designing feasible patterns and scaling to many gaits, with expert-policy distillation and a more versatile representation proposed as future directions.

5 Conclusions

  • The system uses desired foot-contact patterns as a natural-language interface and follows direct and implicit commands in simulation and on physical robots.The conclusion describes the interface as more straightforward and flexible than other design choices.Source: §5 · p. 9

Limitations and Future Work

  • Training patterns must be designed using domain knowledge and trial-and-error to keep them feasible, while expanding gait variety makes training harder due to sample-balancing and data-efficiency trade-offs.The authors identify these constraints as limitations of the approach.Source: Limitations and Future Work · p. 9
  • Future directions include training motion-specific expert policies and distilling them with imitation learning, or changing the representation to specify foot clearance.The proposed representation change could replace binary 0s and 1s with 0s and desired clearance H.Source: Limitations and Future Work · p. 9

A More about Random Pattern Generator

The random pattern generator creates gait templates using a sampled cycle length and contact ratio, gait-specific scaling, and leg-dependent phase shifts. The paper describes four steps and illustrates the process for pacing.

A More about Random Pattern Generator

  • The generator uses four steps to create gait templates, with Figure 7 illustrating the process for pacing.The procedure is based on ranges measured after training in simulation on simple velocity-following tasks.Source: A More about Random Pattern Generator · p. 13, Figure 7 · p. 13
  • Template length T is sampled from 24–28 steps, corresponding to 0.48–0.56 seconds at a 50 Hz control frequency.These settings define the cycle length used by the generator.Source: A More about Random Pattern Generator · p. 13
  • The contact ratio is sampled from 0.5–0.7, determining the number of contact and non-contact entries in each leg’s cycle.The counts are T times the contact ratio for 1s and T times one minus that ratio for 0s.Source: A More about Random Pattern Generator · p. 13
  • For bounding, contact time is shortened to 60% of its sampled value; for pacing, cycle length is halved while contact ratio is unchanged.These adjustments are used because bounding and pacing require shorter foot contact for natural, dynamically feasible motion.Source: A More about Random Pattern Generator · p. 13
  • Leg-specific phase shifts depend on gait: bounding shifts the rear legs by half the contact length, while pacing places front and rear contacts in opposite cycle positions.The generator’s leg shifts require domain knowledge of quadrupedal locomotion.Source: A More about Random Pattern Generator · p. 13

B Reward Design

The controller’s reward is a weighted sum of eight terms spanning task tracking, contact-pattern matching, sim-to-real stability, and natural motion. The terms penalize deviations from desired behavior and undesirable motion or actions.

B Reward Design

  • The reward combines eight weighted terms as a sum, with each term assigned a weight and reward value.The paper groups the terms by task reward, sim-to-real, and natural motion.Source: B Reward Design · p. 13
  • Task rewards track desired forward linear velocity while penalizing lateral velocity (weight 1), use current base-frame yaw velocity (weight −0.5), and penalize foot-contact mismatch (weight −1).Source: B Reward Design · p. 13
  • Sim-to-real terms penalize action-rate changes, roll and pitch angular velocities, and z-axis linear velocity to encourage stable motion.The roll/pitch and z-velocity penalties do not apply to bounding.Source: B Reward Design · pp. 13–14
  • Natural-motion terms penalize body contacts above a force threshold and deviation from the default hip pose.The body-collision term considers eight thighs and calves; the hip-action penalty has weight −0.03.Source: B Reward Design · p. 14

C.1 Control

The system converts position commands to torques using PD control. During simulation training, gait sampling is balanced across five gait types to prevent harder-to-learn gaits from disappearing.

C.1 Control

  • PD control converts positions to torques, using gains k_p = 20 and k_d = 0.5 at a 50 Hz control frequency.These are the system’s base values for the two gains.Source: §C.1 · p. 14

C.2 Gait Sampling

  • The simulation samples a gait at environment resets and again every 150 steps.The gait is assigned to the robot during training.Source: §C.2 · p. 14
  • Sampling keeps the five gait types at approximately equal proportions so harder-to-learn gaits do not die out in favor of easier ones.The authors apply this restriction because some gaits are harder to learn than others.Source: §C.2 · p. 14

C.3 Reinforcement Learning

The controller is trained with PPO using an actor-critic policy. Training adds noise to observations and randomizes the PD gains.

C.3 Reinforcement Learning

  • PPO trains an actor-critic controller; the critic shares the actor’s network architecture but outputs one value instead of 12 and also receives local-frame base velocities.Training runs for 1000 iterations with the hyperparameters used in [42].Source: §C.3 · p. 14
  • Episodes end early when the robot’s body height falls below 0.25 meters, and training takes approximately 15 minutes on one NVIDIA V100 GPU.Source: §C.3 · p. 14

C.4 Domain Randomization

  • Training adds sampled noise to controller observations and applies domain randomization to the PD gains k_p and k_d.Table 3 lists the noise-affected components and their corresponding ranges.Source: §C.4 · p. 14

D Objective Evaluation on Generated Patterns

Generated patterns are evaluated with command-specific checkers based on domain knowledge about quadrupedal locomotion. Table 3 presents the domain randomization settings.

D Objective Evaluation on Generated Patterns

Loading 2306.07580v3…