Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/NVlabs/alpamayo/llms.txt

Use this file to discover all available pages before exploring further.

DiscreteTrajectoryTokenizer converts continuous trajectory data into discrete tokens for use with discrete models, and decodes tokens back into trajectories.

​
Class Definition

class DiscreteTrajectoryTokenizer

​
Constructor

def __init__(
    self,
    action_space_cfg: dict[str, Any],
    dims_min: list[float],
    dims_max: list[float],
    num_bins: int,
    **kwargs: Any,
) -> None
​
action_space_cfg
dict[str, Any]
Configuration dictionary for instantiating the action space using Hydra
​
dims_min
list[float]
Minimum values for each dimension in the action space. Must have length equal to the last dimension of the action space
​
dims_max
list[float]
Maximum values for each dimension in the action space. Must have length equal to the last dimension of the action space
​
num_bins
int
Number of discrete bins to use for quantization
​
**kwargs
Any
Additional keyword arguments (not used)

​
Properties

​
vocab_size

Returns the vocabulary size (number of possible token values). 
@property
def vocab_size(self) -> int
Returns: int - The number of bins, representing tokens from the set {0, 1, ..., vocab_size - 1}

​
Methods

​
encode

Encodes trajectories as discrete tokens. 
def encode(
    self,
    hist_xyz: torch.Tensor,
    hist_rot: torch.Tensor,
    fut_xyz: torch.Tensor,
    fut_rot: torch.Tensor,
    hist_tstamp: torch.Tensor | None = None,
    fut_tstamp: torch.Tensor | None = None,
) -> torch.LongTensor
​
hist_xyz
torch.Tensor
The history xyz coordinates
​
hist_rot
torch.Tensor
The history rotation matrices
​
fut_xyz
torch.Tensor
The future xyz coordinates
​
fut_rot
torch.Tensor
The future rotation matrices
​
hist_tstamp
torch.Tensor | None
default:"None"
The history timestamps (currently unused)
​
fut_tstamp
torch.Tensor | None
default:"None"
The future timestamps (assumed consistent with future trajectory)
Returns: torch.LongTensor with shape (B, num_tokens_per_trajectory) containing the encoded tokens

​
decode

Decodes discrete tokens into future trajectories. 
def decode(
    self,
    hist_xyz: torch.Tensor,
    hist_rot: torch.Tensor,
    tokens: torch.LongTensor,
    hist_tstamp: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]
​
hist_xyz
torch.Tensor
The history xyz coordinates
​
hist_rot
torch.Tensor
The history rotation matrices
​
tokens
torch.LongTensor
The tokens to decode
​
hist_tstamp
torch.Tensor | None
default:"None"
The history timestamps (currently unused)
Returns: Tuple of (fut_xyz, fut_rot, None) where:
  • fut_xyz is the decoded future xyz coordinates
  • fut_rot is the decoded future rotation matrices
  • None indicates future timestamps are not decoded

​
Usage Example

from alpamayo_r1.action_space import DiscreteTrajectoryTokenizer
import torch

# Configure action space
action_space_cfg = {
    "_target_": "alpamayo_r1.action_space.UnicycleAccelCurvatureActionSpace",
    "dt": 0.1,
    "n_waypoints": 64,
}

# Create tokenizer
tokenizer = DiscreteTrajectoryTokenizer(
    action_space_cfg=action_space_cfg,
    dims_min=[-10.0, -0.5],  # Min acceleration and curvature
    dims_max=[10.0, 0.5],     # Max acceleration and curvature
    num_bins=256,
)

print(f"Vocabulary size: {tokenizer.vocab_size}")  # 256

# Encode trajectories
hist_xyz = torch.randn(32, 10, 3)  # Batch of 32, 10 history steps
hist_rot = torch.eye(3).expand(32, 10, 3, 3)
fut_xyz = torch.randn(32, 64, 3)   # 64 future waypoints
fut_rot = torch.eye(3).expand(32, 64, 3, 3)

tokens = tokenizer.encode(hist_xyz, hist_rot, fut_xyz, fut_rot)
print(f"Tokens shape: {tokens.shape}")  # (32, num_tokens)

# Decode tokens back to trajectories
fut_xyz_decoded, fut_rot_decoded, _ = tokenizer.decode(
    hist_xyz, hist_rot, tokens
)
print(f"Decoded trajectory shape: {fut_xyz_decoded.shape}")  # (32, 64, 3)
The tokenizer quantizes continuous action space values into discrete bins. The encoding process:
  1. Converts trajectories to actions using the action space
  2. Normalizes actions to [0, 1] range using dims_min and dims_max
  3. Quantizes to num_bins discrete values
  4. Clamps values to valid range

Build docs developers (and LLMs) love

Get started for freeTalk to us